From 0c882ae846659b10f224f0f98c0c377c21ac0cac Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?=E9=83=AD=E5=BA=86=E6=B3=B0?=
 <12369755+htcloud1@user.noreply.gitee.com>
Date: Mon, 24 Aug 2026 19:23:24 +0800
Subject: [PATCH] =?UTF-8?q?feat:=20=E6=9C=AC=E5=9C=B0=20Java=20=E5=8F=91?=
 =?UTF-8?q?=E7=A5=A8=20OCR=20+=20=E7=AD=96=E5=88=92=E6=96=B9=E6=A1=88?=
 =?UTF-8?q?=E5=A4=9A=E8=A7=92=E8=89=B2=E5=BC=80=E6=94=BE=20+=20H5=20?=
 =?UTF-8?q?=E7=9B=B8=E6=9C=BA=E8=84=B1=E6=95=8F=E4=BF=AE=E5=A4=8D?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

- OCR: 进程内 PaddleOCR ONNX Runtime 识别替代远程 Python 微服务
  (新增 ocr/core/service/config 引擎, 删 OcrClient/OcrConfig, PDF 文本层优先 + QR 快路径)
- 投稿: 项目设计投稿 → 项目策划方案, doctor/executor/sponsor 三角色开放
  (BizProjectPlan 按非 admin/manager 隔离, 复用 Submissions.vue, 设计文件改 OssFileUploader)
- H5: <video> 去掉 autoplay + safePlay 重试, 穿透 uni-app wrapper 找原生 video
- 脱敏: 签到表高斯模糊由横向条带改为全高竖条带 (20%~50% 宽度)
- 域名: ringdoctor.com → risingdoctor.com
---
 .gitignore                                    |   4 +
 .../src/main/resources/application.yml        |  24 +-
 ry-api/ruoyi-business/pom.xml                 |  27 +-
 .../com/ruoyi/business/config/OcrConfig.java  |  23 -
 .../controller/BizProjectPlanController.java  |  22 +-
 .../com/ruoyi/business/ocr/InvoiceResult.java |  12 +-
 .../business/ocr/LocalInvoiceRecognizer.java  |  57 +++
 .../com/ruoyi/business/ocr/OcrClient.java     | 120 -----
 .../java/com/ruoyi/business/ocr/OcrLine.java  |   6 +-
 .../ruoyi/business/ocr/QrDecodeResult.java    |  31 ++
 .../business/ocr/config/OcrProperties.java    |  73 +++
 .../ruoyi/business/ocr/core/CtcDecoder.java   |  53 +++
 .../business/ocr/core/DbPostProcessor.java    | 150 ++++++
 .../ruoyi/business/ocr/core/Dictionary.java   |  56 +++
 .../business/ocr/core/ImageProcessor.java     | 227 +++++++++
 .../ruoyi/business/ocr/core/OcrEngine.java    | 231 ++++++++++
 .../ruoyi/business/ocr/core/PdfProcessor.java |  86 ++++
 .../ruoyi/business/ocr/core/TextDetector.java | 156 +++++++
 .../business/ocr/core/TextRecognizer.java     | 181 ++++++++
 .../ocr/exception/OcrTimeoutException.java    |  17 +
 .../ocr/service/InvoiceExtractor.java         | 434 ++++++++++++++++++
 .../ruoyi/business/ocr/service/QrDecoder.java | 159 +++++++
 .../ocr/service/RecognizeService.java         | 312 +++++++++++++
 .../ruoyi/business/ocr/util/AmountUtils.java  | 296 ++++++++++++
 .../service/impl/InvoiceOcrService.java       |  10 +-
 .../ruoyi/business/sms/AliyunSmsSender.java   |   4 +-
 ry-h5/composables/useCamera.ts                |  22 +-
 ry-h5/pages/panorama/index.uvue               |  28 +-
 ry-h5/pages/signin/index.uvue                 |  30 +-
 ry-h5/utils/camera.ts                         | 173 +++++--
 ry-vue3/src/layout/AdminLayout.vue            |   7 +-
 ry-vue3/src/router/index.js                   |  16 +-
 ry-vue3/src/views/doctor/SubmissionDetail.vue |   6 +-
 ry-vue3/src/views/doctor/SubmissionNew.vue    |  68 +--
 ry-vue3/src/views/doctor/Submissions.vue      |  15 +-
 ry-vue3/src/views/portal/Home.vue             |  11 +-
 36 files changed, 2851 insertions(+), 296 deletions(-)
 delete mode 100644 ry-api/ruoyi-business/src/main/java/com/ruoyi/business/config/OcrConfig.java
 create mode 100644 ry-api/ruoyi-business/src/main/java/com/ruoyi/business/ocr/LocalInvoiceRecognizer.java
 delete mode 100644 ry-api/ruoyi-business/src/main/java/com/ruoyi/business/ocr/OcrClient.java
 create mode 100644 ry-api/ruoyi-business/src/main/java/com/ruoyi/business/ocr/QrDecodeResult.java
 create mode 100644 ry-api/ruoyi-business/src/main/java/com/ruoyi/business/ocr/config/OcrProperties.java
 create mode 100644 ry-api/ruoyi-business/src/main/java/com/ruoyi/business/ocr/core/CtcDecoder.java
 create mode 100644 ry-api/ruoyi-business/src/main/java/com/ruoyi/business/ocr/core/DbPostProcessor.java
 create mode 100644 ry-api/ruoyi-business/src/main/java/com/ruoyi/business/ocr/core/Dictionary.java
 create mode 100644 ry-api/ruoyi-business/src/main/java/com/ruoyi/business/ocr/core/ImageProcessor.java
 create mode 100644 ry-api/ruoyi-business/src/main/java/com/ruoyi/business/ocr/core/OcrEngine.java
 create mode 100644 ry-api/ruoyi-business/src/main/java/com/ruoyi/business/ocr/core/PdfProcessor.java
 create mode 100644 ry-api/ruoyi-business/src/main/java/com/ruoyi/business/ocr/core/TextDetector.java
 create mode 100644 ry-api/ruoyi-business/src/main/java/com/ruoyi/business/ocr/core/TextRecognizer.java
 create mode 100644 ry-api/ruoyi-business/src/main/java/com/ruoyi/business/ocr/exception/OcrTimeoutException.java
 create mode 100644 ry-api/ruoyi-business/src/main/java/com/ruoyi/business/ocr/service/InvoiceExtractor.java
 create mode 100644 ry-api/ruoyi-business/src/main/java/com/ruoyi/business/ocr/service/QrDecoder.java
 create mode 100644 ry-api/ruoyi-business/src/main/java/com/ruoyi/business/ocr/service/RecognizeService.java
 create mode 100644 ry-api/ruoyi-business/src/main/java/com/ruoyi/business/ocr/util/AmountUtils.java

diff --git a/.gitignore b/.gitignore
index 4d5507e..dab9666 100644
--- a/.gitignore
+++ b/.gitignore
@@ -84,3 +84,7 @@ ry8080/
 ry8080.zip
 ry-console/
 ry-vue2/  # 旧 vue 控制台
+
+# ===== OCR 模型 (大二进制, 不进版本库, 部署时放 jar 同目录 models/) =====
+ry-api/models/
+ry-ocr-java/src/main/resources/models/
diff --git a/ry-api/ruoyi-admin/src/main/resources/application.yml b/ry-api/ruoyi-admin/src/main/resources/application.yml
index df90b96..82087de 100644
--- a/ry-api/ruoyi-admin/src/main/resources/application.yml
+++ b/ry-api/ruoyi-admin/src/main/resources/application.yml
@@ -31,15 +31,31 @@ ruoyi:
     signName: 北京仙仁掌医学科技发展
     template: SMS_321560247
     esignTemplate: SMS_492460505
-    esignBaseUrl: https://ringdoctor.com/hg
+    esignBaseUrl: https://risingdoctor.com/hg
     endpoint: dysmsapi.aliyuncs.com
     regionId: cn-hangzhou
-  # 发票 OCR (ry-ocr 微服务, PaddleOCR + FastAPI, 默认 http://127.0.0.1:8801)
+  # 发票 OCR (本地 Java 识别, PaddleOCR ONNX Runtime, 替代原 ry-ocr Python 微服务)
   ocr:
-    base-url: http://127.0.0.1:8801
+    version: 0.1.0
+    # 模型目录: 绝对路径 → 直读; 相对路径 → 相对进程工作目录 (java -jar 启动目录 = jar 同目录, 部署时 models/ 与 jar 同级)
+    models-dir: models
+    # 模型版本: v5_mobile (H=48, 20MB, 推荐) / v5_server (90MB+) / v4_mobile (15MB, 最快)
+    model-version: v5_mobile
+    # 单页 OCR 超时 / 整流程超时 (秒)
+    page-timeout-s: 15
+    total-timeout-s: 60
+    # QR 命中后是否继续跑全量 OCR (false = 快路径只返回 QR 3 字段)
+    qr-full-ocr: true
+    lang: ch
+    # PDF 优先抽内嵌文本层 (电子发票秒出, 扫描件自动回退 ONNX)
+    use-pdf-text-first: true
+    pdf-text-min-chars: 30
+    upload:
+      max-mb: 20
+      pdf-dpi: 150
   # 扫码拍照 (ry-h5 相机网页地址, 前端二维码目标 URL)
   camera:
-    base-url: https://ringdoctor.com/camera/
+    base-url: https://risingdoctor.com/camera/
 
 # 开发环境配置
 server:
diff --git a/ry-api/ruoyi-business/pom.xml b/ry-api/ruoyi-business/pom.xml
index 7619979..4859ee1 100644
--- a/ry-api/ruoyi-business/pom.xml
+++ b/ry-api/ruoyi-business/pom.xml
@@ -51,7 +51,7 @@
             <artifactId>html2pdf</artifactId>
             <version>3.0.2</version>
         </dependency>
-        <!-- ry-ocr Java 调用客户端依赖: hutool-http / hutool-json / hutool-core -->
+        <!-- hutool (本地 OCR 下载 / 文件 / JSON 工具) -->
         <dependency>
             <groupId>cn.hutool</groupId>
             <artifactId>hutool-http</artifactId>
@@ -67,7 +67,30 @@
             <artifactId>hutool-core</artifactId>
             <version>5.8.27</version>
         </dependency>
-        <!-- Lombok (OcrClient 等用了 @Slf4j) -->
+        <!-- ===== 本地发票 OCR (PaddleOCR ONNX Runtime) ===== -->
+        <dependency>
+            <groupId>com.microsoft.onnxruntime</groupId>
+            <artifactId>onnxruntime</artifactId>
+            <version>1.20.0</version>
+        </dependency>
+        <!-- PDF 渲染 + 文本层抽取 -->
+        <dependency>
+            <groupId>org.apache.pdfbox</groupId>
+            <artifactId>pdfbox</artifactId>
+            <version>2.0.31</version>
+        </dependency>
+        <!-- 二维码解码 -->
+        <dependency>
+            <groupId>com.google.zxing</groupId>
+            <artifactId>core</artifactId>
+            <version>3.5.3</version>
+        </dependency>
+        <dependency>
+            <groupId>com.google.zxing</groupId>
+            <artifactId>javase</artifactId>
+            <version>3.5.3</version>
+        </dependency>
+        <!-- Lombok (OCR 引擎等用了 @Slf4j) -->
         <dependency>
             <groupId>org.projectlombok</groupId>
             <artifactId>lombok</artifactId>
diff --git a/ry-api/ruoyi-business/src/main/java/com/ruoyi/business/config/OcrConfig.java b/ry-api/ruoyi-business/src/main/java/com/ruoyi/business/config/OcrConfig.java
deleted file mode 100644
index 473869a..0000000
--- a/ry-api/ruoyi-business/src/main/java/com/ruoyi/business/config/OcrConfig.java
+++ /dev/null
@@ -1,23 +0,0 @@
-package com.ruoyi.business.config;
-
-import org.springframework.beans.factory.annotation.Value;
-import org.springframework.context.annotation.Bean;
-import org.springframework.context.annotation.Configuration;
-import com.ruoyi.business.ocr.OcrClient;
-
-/**
- * ry-ocr 微服务集成配置
- * <p>
- * yml 配置: ruoyi.ocr.base-url (默认 http://127.0.0.1:8801)
- */
-@Configuration
-public class OcrConfig {
-
-    @Value("${ruoyi.ocr.base-url:http://127.0.0.1:8801}")
-    private String ocrBaseUrl;
-
-    @Bean
-    public OcrClient ocrClient() {
-        return new OcrClient(ocrBaseUrl);
-    }
-}
\ No newline at end of file
diff --git a/ry-api/ruoyi-business/src/main/java/com/ruoyi/business/controller/BizProjectPlanController.java b/ry-api/ruoyi-business/src/main/java/com/ruoyi/business/controller/BizProjectPlanController.java
index 980912a..6813cc8 100644
--- a/ry-api/ruoyi-business/src/main/java/com/ruoyi/business/controller/BizProjectPlanController.java
+++ b/ry-api/ruoyi-business/src/main/java/com/ruoyi/business/controller/BizProjectPlanController.java
@@ -16,8 +16,8 @@ import com.ruoyi.business.service.IBizProjectPlanService;
  * 项目策划方案Controller
  *
  * 角色权限:
- *   - doctor: 只看自己投的稿 (submitter_id = 当前用户); 新建/编辑强制 submitter_id 写自己, status 默认 '0'
- *   - manager / sponsor / admin: 全部可见, 不强制 submitter
+ *   - 投稿角色 (doctor/executor/sponsor): 只看自己投的稿 (submitter_id = 当前用户); 新建/编辑强制 submitter_id 写自己, status 默认 '0'
+ *   - 管理角色 (admin/manager): 全部可见, 不强制 submitter (用于审核/结算)
  */
 @RestController
 @RequestMapping("/business/projectPlan")
@@ -28,9 +28,9 @@ public class BizProjectPlanController extends BaseController
     @GetMapping("/list")
     public TableDataInfo list(BizProjectPlan BizProjectPlan)
     {
-        // 医生角色: 后端兜底只查自己投的稿
+        // 投稿角色: 后端兜底只查自己投的稿
         String roleType = SecurityUtils.getLoginUser().getUser().getRoleType();
-        if ("doctor".equals(roleType)) {
+        if (isSubmitterRole(roleType)) {
             BizProjectPlan.setSubmitterId(SecurityUtils.getUserId());
         }
         startPage();
@@ -46,9 +46,9 @@ public class BizProjectPlanController extends BaseController
     @PostMapping
     public AjaxResult add(@RequestBody BizProjectPlan BizProjectPlan)
     {
-        // 医生角色: 强制 submitter_id 写自己 + 状态兜底 '0' (未提交), 防止绕过
+        // 投稿角色: 强制 submitter_id 写自己 + 状态兜底 '0' (未提交), 防止绕过
         String roleType = SecurityUtils.getLoginUser().getUser().getRoleType();
-        if ("doctor".equals(roleType)) {
+        if (isSubmitterRole(roleType)) {
             BizProjectPlan.setSubmitterId(SecurityUtils.getUserId());
             if (BizProjectPlan.getStatus() == null || BizProjectPlan.getStatus().isEmpty()) {
                 BizProjectPlan.setStatus("0");
@@ -60,9 +60,9 @@ public class BizProjectPlanController extends BaseController
     @PutMapping
     public AjaxResult edit(@RequestBody BizProjectPlan BizProjectPlan)
     {
-        // 医生角色: 修改时也强制覆盖 submitter_id, 防止越权篡改
+        // 投稿角色: 修改时也强制覆盖 submitter_id, 防止越权篡改
         String roleType = SecurityUtils.getLoginUser().getUser().getRoleType();
-        if ("doctor".equals(roleType)) {
+        if (isSubmitterRole(roleType)) {
             BizProjectPlan.setSubmitterId(SecurityUtils.getUserId());
         }
         return toAjax(BizProjectPlanService.updateByPrimaryKey(BizProjectPlan));
@@ -73,4 +73,10 @@ public class BizProjectPlanController extends BaseController
     {
         return toAjax(BizProjectPlanService.deleteByPrimaryKeys(ids));
     }
+
+    /** 投稿角色 = 非 admin/manager (doctor/executor/sponsor): 只看/只写自己的稿 */
+    private boolean isSubmitterRole(String roleType)
+    {
+        return !("admin".equals(roleType) || "manager".equals(roleType));
+    }
 }
\ No newline at end of file
diff --git a/ry-api/ruoyi-business/src/main/java/com/ruoyi/business/ocr/InvoiceResult.java b/ry-api/ruoyi-business/src/main/java/com/ruoyi/business/ocr/InvoiceResult.java
index ea55cfc..8d90c9e 100644
--- a/ry-api/ruoyi-business/src/main/java/com/ruoyi/business/ocr/InvoiceResult.java
+++ b/ry-api/ruoyi-business/src/main/java/com/ruoyi/business/ocr/InvoiceResult.java
@@ -8,11 +8,21 @@ import java.util.List;
 @Data
 public class InvoiceResult {
     private Boolean success;
-    private String rawText;
+    /** 是否被判定为发票 (false = 非发票图片) */
+    private Boolean isInvoice = true;
+    private String rawText = "";
     private String engine;
     private Integer pageCount;
     private Integer elapsedMs;
     private String error;
+    /** 错误码: not_invoice / timeout / unsupported / ocr_failed / process_failed */
+    private String errorCode;
+    /** 是否从 QR 取到了核心字段 */
+    private Boolean fromQr = false;
+    /** 二维码原始文本 (排查用) */
+    private String qrRaw;
+    /** 二维码识别失败原因 (no_qr / bad_format) */
+    private String qrError;
     private InvoiceFields fields;
     private List<OcrLine> lines;
 }
diff --git a/ry-api/ruoyi-business/src/main/java/com/ruoyi/business/ocr/LocalInvoiceRecognizer.java b/ry-api/ruoyi-business/src/main/java/com/ruoyi/business/ocr/LocalInvoiceRecognizer.java
new file mode 100644
index 0000000..f5565ea
--- /dev/null
+++ b/ry-api/ruoyi-business/src/main/java/com/ruoyi/business/ocr/LocalInvoiceRecognizer.java
@@ -0,0 +1,57 @@
+package com.ruoyi.business.ocr;
+
+import cn.hutool.core.io.FileUtil;
+import cn.hutool.http.HttpUtil;
+import com.ruoyi.business.ocr.service.RecognizeService;
+import lombok.RequiredArgsConstructor;
+import lombok.extern.slf4j.Slf4j;
+import org.springframework.stereotype.Service;
+
+import java.io.File;
+
+/**
+ * 本地发票识别门面 — 替代原 OcrClient (HTTP 调 Python ry-ocr 微服务).
+ * <p>
+ * 直接委托 {@link RecognizeService} 在进程内跑 PaddleOCR ONNX 推理,
+ * 对外暴露与原 OcrClient 相同的两个方法, 调用方 (InvoiceOcrService) 无需改动.
+ */
+@Slf4j
+@Service
+@RequiredArgsConstructor
+public class LocalInvoiceRecognizer {
+
+    private final RecognizeService recognizeService;
+
+    /** 识别发票 (图片或 PDF) */
+    public InvoiceResult recognize(File file) {
+        byte[] content = FileUtil.readBytes(file);
+        return recognizeService.recognizeFile(file.getName(), content);
+    }
+
+    /**
+     * 从 URL 识别发票: 下载 OSS URL 到临时文件 → 本地识别 → 清理临时文件.
+     */
+    public InvoiceResult recognizeByUrl(String url) {
+        if (url == null || url.isEmpty()) {
+            throw new IllegalArgumentException("ossUrl 不能为空");
+        }
+        File tmpDir = new File(System.getProperty("java.io.tmpdir"), "ry-ocr");
+        if (!tmpDir.exists() && !tmpDir.mkdirs()) {
+            throw new RuntimeException("无法创建临时目录: " + tmpDir.getAbsolutePath());
+        }
+        String name = url.substring(url.lastIndexOf('/') + 1);
+        if (name.indexOf('?') >= 0) name = name.substring(0, name.indexOf('?'));
+        if (name.indexOf('.') < 0) name = name + ".png";
+        File tmp = new File(tmpDir, System.currentTimeMillis() + "_" + name);
+        try {
+            long size = HttpUtil.downloadFile(url, tmp);
+            if (size <= 0) {
+                throw new RuntimeException("OSS 文件下载失败或为空: " + url);
+            }
+            log.info("OCR 下载: url={} size={}B tmp={}", url, size, tmp.getAbsolutePath());
+            return recognize(tmp);
+        } finally {
+            FileUtil.del(tmp);
+        }
+    }
+}
diff --git a/ry-api/ruoyi-business/src/main/java/com/ruoyi/business/ocr/OcrClient.java b/ry-api/ruoyi-business/src/main/java/com/ruoyi/business/ocr/OcrClient.java
deleted file mode 100644
index 72558f4..0000000
--- a/ry-api/ruoyi-business/src/main/java/com/ruoyi/business/ocr/OcrClient.java
+++ /dev/null
@@ -1,120 +0,0 @@
-package com.ruoyi.business.ocr;
-
-import cn.hutool.core.io.FileUtil;
-import cn.hutool.http.HttpRequest;
-import cn.hutool.http.HttpResponse;
-import cn.hutool.http.HttpUtil;
-import cn.hutool.json.JSONObject;
-import cn.hutool.json.JSONUtil;
-import lombok.extern.slf4j.Slf4j;
-
-import java.io.File;
-
-/**
- * ry-ocr Java 调用客户端
- *
- * 依赖：hutool-http, hutool-json, hutool-core, lombok
- *
- * 用法：
- *   OcrClient client = new OcrClient("http://127.0.0.1:8801");
- *   InvoiceResult r = client.recognize(new File("d:/发票.pdf"));
- *   InvoiceResult r2 = client.recognizeByUrl("https://oss.example.com/xxx.png");
- *   System.out.println(r.getFields().getAmount());
- */
-@Slf4j
-public class OcrClient {
-
-    private final String baseUrl;
-
-    public OcrClient(String baseUrl) {
-        this.baseUrl = baseUrl.endsWith("/") ? baseUrl.substring(0, baseUrl.length() - 1) : baseUrl;
-    }
-
-    /** 健康检查 */
-    public boolean ping() {
-        try (HttpResponse resp = HttpRequest.get(baseUrl + "/health").timeout(3000).execute()) {
-            return resp.getStatus() == 200 && "ok".equals(JSONUtil.parseObj(resp.body()).getStr("status"));
-        } catch (Exception e) {
-            log.warn("ocr ping failed: {}", e.getMessage());
-            return false;
-        }
-    }
-
-    /** 识别发票（图片或 PDF） */
-    public InvoiceResult recognize(File file) {
-        try (HttpResponse resp = HttpRequest.post(baseUrl + "/recognize/invoice")
-                .form("file", file)
-                .timeout(60_000)
-                .execute()) {
-
-            String body = resp.body();
-            JSONObject json = JSONUtil.parseObj(body);
-            if (resp.getStatus() != 200) {
-                throw new RuntimeException("OCR 调用失败: " + resp.getStatus() + " " + body);
-            }
-            return parse(json);
-        }
-    }
-
-    /**
-     * 从 URL 识别发票: 后端下载 OSS URL 到临时文件 → recognize → 清理临时文件.
-     * 临时文件目录: System.getProperty("java.io.tmpdir")/ry-ocr/
-     *
-     * @param url OSS 可访问 URL
-     * @return 识别结果
-     */
-    public InvoiceResult recognizeByUrl(String url) {
-        if (url == null || url.isEmpty()) {
-            throw new IllegalArgumentException("ossUrl 不能为空");
-        }
-        File tmpDir = new File(System.getProperty("java.io.tmpdir"), "ry-ocr");
-        if (!tmpDir.exists() && !tmpDir.mkdirs()) {
-            throw new RuntimeException("无法创建临时目录: " + tmpDir.getAbsolutePath());
-        }
-        // 从 URL 截取文件名, 保留后缀 (用于 ry-ocr 推断图片/PDF)
-        String name = url.substring(url.lastIndexOf('/') + 1);
-        if (name.indexOf('?') >= 0) name = name.substring(0, name.indexOf('?'));
-        if (name.indexOf('.') < 0) name = name + ".png";
-        File tmp = new File(tmpDir, System.currentTimeMillis() + "_" + name);
-        try {
-            long size = HttpUtil.downloadFile(url, tmp);
-            if (size <= 0) {
-                throw new RuntimeException("OSS 文件下载失败或为空: " + url);
-            }
-            log.info("OCR 下载: url={} size={}B tmp={}", url, size, tmp.getAbsolutePath());
-            return recognize(tmp);
-        } finally {
-            FileUtil.del(tmp);
-        }
-    }
-
-    private InvoiceResult parse(JSONObject json) {
-        InvoiceResult r = new InvoiceResult();
-        r.setSuccess(json.getBool("success", false));
-        r.setRawText(json.getStr("rawText", ""));
-        r.setEngine(json.getStr("engine", ""));
-        r.setPageCount(json.getInt("pageCount", 1));
-        r.setElapsedMs(json.getInt("elapsedMs", 0));
-        r.setError(json.getStr("error"));
-
-        JSONObject f = json.getJSONObject("fields");
-        if (f != null) {
-            InvoiceFields fields = new InvoiceFields();
-            fields.setInvoiceType(f.getStr("invoiceType"));
-            fields.setInvoiceNo(f.getStr("invoiceNo"));
-            fields.setInvoiceCode(f.getStr("invoiceCode"));
-            fields.setInvoiceDate(f.getStr("invoiceDate"));
-            fields.setAmount(f.getDouble("amount"));
-            fields.setAmountCn(f.getStr("amountCn"));
-            fields.setAmountPretax(f.getDouble("amount_pretax"));
-            fields.setTaxAmount(f.getDouble("taxAmount"));
-            fields.setSellerName(f.getStr("sellerName"));
-            fields.setSellerTaxNo(f.getStr("sellerTaxNo"));
-            fields.setBuyerName(f.getStr("buyerName"));
-            fields.setBuyerTaxNo(f.getStr("buyerTaxNo"));
-            fields.setAmountMatch(f.getBool("amountMatch"));
-            r.setFields(fields);
-        }
-        return r;
-    }
-}
diff --git a/ry-api/ruoyi-business/src/main/java/com/ruoyi/business/ocr/OcrLine.java b/ry-api/ruoyi-business/src/main/java/com/ruoyi/business/ocr/OcrLine.java
index e4c3ccb..4919939 100644
--- a/ry-api/ruoyi-business/src/main/java/com/ruoyi/business/ocr/OcrLine.java
+++ b/ry-api/ruoyi-business/src/main/java/com/ruoyi/business/ocr/OcrLine.java
@@ -1,13 +1,17 @@
 package com.ruoyi.business.ocr;
 
+import lombok.AllArgsConstructor;
 import lombok.Data;
+import lombok.NoArgsConstructor;
 
 import java.util.List;
 
 /** 单行 OCR 识别结果 */
 @Data
+@NoArgsConstructor
+@AllArgsConstructor
 public class OcrLine {
     private String text;
     private Double confidence;
-    private List<List<Double>> box;
+    private List<List<Float>> box;
 }
diff --git a/ry-api/ruoyi-business/src/main/java/com/ruoyi/business/ocr/QrDecodeResult.java b/ry-api/ruoyi-business/src/main/java/com/ruoyi/business/ocr/QrDecodeResult.java
new file mode 100644
index 0000000..fcb6116
--- /dev/null
+++ b/ry-api/ruoyi-business/src/main/java/com/ruoyi/business/ocr/QrDecodeResult.java
@@ -0,0 +1,31 @@
+package com.ruoyi.business.ocr;
+
+import lombok.AllArgsConstructor;
+import lombok.Data;
+import lombok.NoArgsConstructor;
+
+/**
+ * 二维码识别结果 — 对齐 Python app.services.qr_decoder.QRDecodeResult
+ */
+@Data
+@NoArgsConstructor
+@AllArgsConstructor
+public class QrDecodeResult {
+
+    /** 发票号码 */
+    private String invoiceNo;
+
+    /** 金额 (小写) */
+    private Double amount;
+
+    /** 开票日期 YYYY-MM-DD */
+    private String invoiceDate;
+
+    /** 二维码原始文本 */
+    private String raw = "";
+
+    /** 是否有任一关键字段解出 */
+    public boolean hasAnyField() {
+        return invoiceNo != null || amount != null || invoiceDate != null;
+    }
+}
diff --git a/ry-api/ruoyi-business/src/main/java/com/ruoyi/business/ocr/config/OcrProperties.java b/ry-api/ruoyi-business/src/main/java/com/ruoyi/business/ocr/config/OcrProperties.java
new file mode 100644
index 0000000..f18853f
--- /dev/null
+++ b/ry-api/ruoyi-business/src/main/java/com/ruoyi/business/ocr/config/OcrProperties.java
@@ -0,0 +1,73 @@
+package com.ruoyi.business.ocr.config;
+
+import lombok.Data;
+import org.springframework.boot.context.properties.ConfigurationProperties;
+import org.springframework.stereotype.Component;
+
+import java.util.ArrayList;
+import java.util.List;
+
+/**
+ * 本地发票 OCR 配置 (移植自 ry-ocr-java, 对应 Python app/config.py + .env)
+ * <p>
+ * 绑定 application.yml 中 ruoyi.ocr.* 配置段.
+ */
+@Data
+@Component
+@ConfigurationProperties(prefix = "ruoyi.ocr")
+public class OcrProperties {
+
+    /** 服务版本号 (对应 Python __version__) */
+    private String version = "0.1.0";
+
+    private final Upload upload = new Upload();
+    private final Ocr ocr = new Ocr();
+
+    /**
+     * 路径白名单: 空字符串/空列表 = 禁用 by-path 接口
+     * <p>
+     * 支持: yml 数组 ["E:\\a", "D:\\b"] 或单字符串 "E:\\a;D:\\b" (Windows 分号分隔)
+     */
+    private List<String> allowedDirs = new ArrayList<>();
+
+    @Data
+    public static class Upload {
+        /** 单文件最大 MB, 超限返回 HTTP 413 */
+        private int maxMb = 20;
+        /** PDF 转图片 DPI */
+        private int pdfDpi = 150;
+    }
+
+    @Data
+    public static class Ocr {
+        /** 单页 OCR 超时 (秒) */
+        private int pageTimeoutS = 15;
+        /** 整流程 OCR 超时 (秒) */
+        private int totalTimeoutS = 60;
+        /** QR 命中后是否继续跑全量 OCR + 字段抽取 (false = 快路径, 只返回 QR 3 字段) */
+        private boolean qrFullOcr = true;
+        /** OCR 语言: ch (简中) / en / chinese_cht */
+        private String lang = "ch";
+        /** 模型目录: 相对路径 → classpath:models/, 绝对路径 → 直读 */
+        private String modelsDir = "models";
+        /** PDF 文件优先抽内嵌文本层 (pdftotext 等价) — 抽到非空文本则跳过 ONNX.
+         *  适用电子发票 / 数电票 PDF (含真实文本); 扫描件 PDF 会回退到 ONNX. */
+        private boolean usePdfTextFirst = true;
+        /** 内嵌文本字符数低于此值视为无效, 回退到 ONNX */
+        private int pdfTextMinChars = 30;
+        /** 模型版本 (决定 CRNN 输入高度):
+         *  <ul>
+         *    <li>v5_server — H=48, maxW=320 (PaddleOCR v5 server 多语言模型, 90MB+ 大, 精度最高)</li>
+         *    <li>v5_mobile — H=48, maxW=320 (PaddleOCR v5 mobile 多语言模型, 20MB, 推荐折中)</li>
+         *    <li>v4_mobile — H=32, maxW=320 (PaddleOCR v4 mobile 中文模型, 15MB, 最快, 中文精度略低)</li>
+         *  </ul>
+         *  切换时改 models-dir + 本字段 + 重启即可. */
+        private String modelVersion = "v5_mobile";
+        /** CRNN 输入高度 (覆盖 modelVersion 默认值). 高级用户用, 一般不动. */
+        private Integer recHeight = null;
+        /** CRNN 最大宽度 (覆盖 modelVersion 默认值) */
+        private Integer recMaxWidth = null;
+        /** DB 检测最长边限制 (覆盖 modelVersion 默认值) */
+        private Integer detMaxSide = null;
+    }
+}
diff --git a/ry-api/ruoyi-business/src/main/java/com/ruoyi/business/ocr/core/CtcDecoder.java b/ry-api/ruoyi-business/src/main/java/com/ruoyi/business/ocr/core/CtcDecoder.java
new file mode 100644
index 0000000..7104026
--- /dev/null
+++ b/ry-api/ruoyi-business/src/main/java/com/ruoyi/business/ocr/core/CtcDecoder.java
@@ -0,0 +1,53 @@
+package com.ruoyi.business.ocr.core;
+
+/**
+ * CTC greedy decoder: argmax → 去连续重复 → 去 blank (idx 0) → 字典查表.
+ */
+public final class CtcDecoder {
+
+    private CtcDecoder() {}
+
+    /**
+     * @param logits [T, N]
+     * @param dict   字典 (idx 0 = blank)
+     * @return RecognizedText
+     */
+    public static TextRecognizer.RecognizedText decode(float[][] logits, Dictionary dict) {
+        int t = logits.length;
+        if (t == 0) return new TextRecognizer.RecognizedText("", 0.0);
+
+        StringBuilder sb = new StringBuilder();
+        int lastIdx = -1;
+        double confSum = 0;
+        int confCount = 0;
+
+        for (int i = 0; i < t; i++) {
+            int bestIdx = 0;
+            float bestVal = Float.NEGATIVE_INFINITY;
+            for (int j = 0; j < logits[i].length; j++) {
+                if (logits[i][j] > bestVal) {
+                    bestVal = logits[i][j];
+                    bestIdx = j;
+                }
+            }
+            if (bestIdx != 0 && bestIdx != lastIdx) {
+                // 跳过 blank (0), 跳过与上一次相同的 (CTC 合并)
+                if (bestIdx < dict.size()) {
+                    sb.append(dict.getCharacters().get(bestIdx));
+                }
+                // softmax → exp / sum
+                double sumExp = 0;
+                for (int j = 0; j < logits[i].length; j++) {
+                    sumExp += Math.exp(logits[i][j] - bestVal);
+                }
+                double prob = 1.0 / sumExp;
+                confSum += prob;
+                confCount++;
+            }
+            lastIdx = bestIdx;
+        }
+
+        double conf = confCount == 0 ? 0 : confSum / confCount;
+        return new TextRecognizer.RecognizedText(sb.toString(), conf);
+    }
+}
diff --git a/ry-api/ruoyi-business/src/main/java/com/ruoyi/business/ocr/core/DbPostProcessor.java b/ry-api/ruoyi-business/src/main/java/com/ruoyi/business/ocr/core/DbPostProcessor.java
new file mode 100644
index 0000000..00d89a6
--- /dev/null
+++ b/ry-api/ruoyi-business/src/main/java/com/ruoyi/business/ocr/core/DbPostProcessor.java
@@ -0,0 +1,150 @@
+package com.ruoyi.business.ocr.core;
+
+import java.awt.geom.Path2D;
+import java.awt.geom.PathIterator;
+import java.util.ArrayList;
+import java.util.List;
+
+/**
+ * DB (Differentiable Binarization) 后处理 — 从概率图提取文本框 polygon.
+ * <p>
+ * 复刻 Python PaddleOCR 的 db_post_process / boxes_from_bitmap.
+ */
+public final class DbPostProcessor {
+
+    private DbPostProcessor() {}
+
+    /**
+     * @param prob        [H, W] 概率图
+     * @param resizeH/W   概率图对应的输入图尺寸
+     * @param origH/W     原图尺寸 (用于映射回原图坐标)
+     */
+    public static List<List<Float>> postProcess(float[][] prob, int resizeH, int resizeW,
+                                                int origH, int origW,
+                                                float dbThresh, float boxThresh,
+                                                float unclipRatio) {
+        // 1. 二值化 + 膨胀 (这里简化为阈值 + 内置 unclipRatio 计算 box)
+        //    生产实现需要 findContours, 这里用简化: 标记连通分量 + bounding box + unclip
+        boolean[][] mask = new boolean[resizeH][resizeW];
+        for (int y = 0; y < resizeH; y++) {
+            for (int x = 0; x < resizeW; x++) {
+                mask[y][x] = prob[y][x] >= dbThresh;
+            }
+        }
+        // 简易膨胀 (3x3)
+        boolean[][] dilated = dilate(mask, resizeH, resizeW);
+        // 简易 8 邻接连通分量
+        List<int[][]> components = connectedComponents(dilated, resizeH, resizeW);
+        // 过滤 + unclip
+        float scaleX = (float) origW / resizeW;
+        float scaleY = (float) origH / resizeH;
+        List<List<Float>> boxes = new ArrayList<>();
+        for (int[][] component : components) {
+            int minX = component[0][0], minY = component[0][1];
+            int maxX = component[0][0], maxY = component[0][1];
+            int area = 0;
+            for (int[] p : component) {
+                if (p[0] < minX) minX = p[0];
+                if (p[0] > maxX) maxX = p[0];
+                if (p[1] < minY) minY = p[1];
+                if (p[1] > maxY) maxY = p[1];
+                area++;
+            }
+            // box_thresh 过滤: 用平均 prob 二次过滤
+            if (area < 3) continue;
+            float meanProb = 0;
+            for (int[] p : component) {
+                meanProb += prob[p[1]][p[0]];
+            }
+            meanProb /= area;
+            if (meanProb < boxThresh) continue;
+
+            // unclip: 扩展 box (简化为固定比例放大)
+            int w = maxX - minX + 1, h = maxY - minY + 1;
+            int dx = (int) (w * (unclipRatio - 1) / 2);
+            int dy = (int) (h * (unclipRatio - 1) / 2);
+            minX = Math.max(0, minX - dx);
+            minY = Math.max(0, minY - dy);
+            maxX = Math.min(resizeW - 1, maxX + dx);
+            maxY = Math.min(resizeH - 1, maxY + dy);
+
+            List<Float> box = new ArrayList<>();
+            float[][] corners = {
+                    {minX * scaleX, minY * scaleY},
+                    {maxX * scaleX, minY * scaleY},
+                    {maxX * scaleX, maxY * scaleY},
+                    {minX * scaleX, maxY * scaleY}
+            };
+            for (float[] c : corners) {
+                box.add(c[0]);
+                box.add(c[1]);
+            }
+            boxes.add(box);
+        }
+        // 按 y 排序 (从上到下)
+        boxes.sort((a, b) -> Float.compare(a.get(1), b.get(1)));
+        return boxes;
+    }
+
+    private static boolean[][] dilate(boolean[][] src, int h, int w) {
+        boolean[][] dst = new boolean[h][w];
+        for (int y = 0; y < h; y++) {
+            for (int x = 0; x < w; x++) {
+                boolean any = false;
+                for (int dy = -1; dy <= 1 && !any; dy++) {
+                    for (int dx = -1; dx <= 1 && !any; dx++) {
+                        int ny = y + dy, nx = x + dx;
+                        if (ny >= 0 && ny < h && nx >= 0 && nx < w && src[ny][nx]) {
+                            any = true;
+                        }
+                    }
+                }
+                dst[y][x] = any;
+            }
+        }
+        return dst;
+    }
+
+    private static List<int[][]> connectedComponents(boolean[][] mask, int h, int w) {
+        boolean[][] visited = new boolean[h][w];
+        List<int[][]> result = new ArrayList<>();
+        int[] dx = {-1, 0, 1, -1, 1, -1, 0, 1};
+        int[] dy = {-1, -1, -1, 0, 0, 1, 1, 1};
+        for (int y = 0; y < h; y++) {
+            for (int x = 0; x < w; x++) {
+                if (!mask[y][x] || visited[y][x]) continue;
+                List<int[]> comp = new ArrayList<>();
+                java.util.Deque<int[]> stack = new java.util.ArrayDeque<>();
+                stack.push(new int[]{x, y});
+                visited[y][x] = true;
+                while (!stack.isEmpty()) {
+                    int[] p = stack.pop();
+                    comp.add(p);
+                    for (int i = 0; i < 8; i++) {
+                        int nx = p[0] + dx[i], ny = p[1] + dy[i];
+                        if (nx >= 0 && nx < w && ny >= 0 && ny < h && mask[ny][nx] && !visited[ny][nx]) {
+                            visited[ny][nx] = true;
+                            stack.push(new int[]{nx, ny});
+                        }
+                    }
+                }
+                result.add(comp.toArray(new int[0][]));
+            }
+        }
+        return result;
+    }
+
+    // 工具: polygon → bounding box
+    public static int[] bbox(List<Float> box) {
+        float minX = Float.MAX_VALUE, minY = Float.MAX_VALUE;
+        float maxX = -Float.MAX_VALUE, maxY = -Float.MAX_VALUE;
+        for (int i = 0; i < box.size(); i += 2) {
+            float x = box.get(i), y = box.get(i + 1);
+            if (x < minX) minX = x;
+            if (x > maxX) maxX = x;
+            if (y < minY) minY = y;
+            if (y > maxY) maxY = y;
+        }
+        return new int[]{Math.round(minX), Math.round(minY), Math.round(maxX), Math.round(maxY)};
+    }
+}
diff --git a/ry-api/ruoyi-business/src/main/java/com/ruoyi/business/ocr/core/Dictionary.java b/ry-api/ruoyi-business/src/main/java/com/ruoyi/business/ocr/core/Dictionary.java
new file mode 100644
index 0000000..a3ca660
--- /dev/null
+++ b/ry-api/ruoyi-business/src/main/java/com/ruoyi/business/ocr/core/Dictionary.java
@@ -0,0 +1,56 @@
+package com.ruoyi.business.ocr.core;
+
+import lombok.Getter;
+
+import java.io.BufferedReader;
+import java.io.IOException;
+import java.nio.charset.StandardCharsets;
+import java.nio.file.Files;
+import java.nio.file.Path;
+import java.util.ArrayList;
+import java.util.List;
+
+/**
+ * 中文字典加载 — 加载 PaddleOCR ppocr_keys_v1.txt
+ * <p>
+ * 模型输出约定: PP-OCRv5 multilingual 的输出维度 = len(dict_file_lines) + 2, 其中
+ * <ul>
+ *   <li>idx 0 = CTC blank (不查表, 由解码器跳过)</li>
+ *   <li>idx 1 = " " (半角空格, 多语言模型中英文混排用, 文件中无此行)</li>
+ *   <li>idx 2..N+1 = 字典字符 (来自 ppocr_keys_v1.txt 的每一行)</li>
+ * </ul>
+ * 因此本类在加载时**主动在 idx 0 插入空串占位, idx 1 插入半角空格**, 使
+ * dict.characters[2] 对应文件第 1 行.
+ * <p>
+ * ⚠️ 此约定根据 monkt/paddleocr-onnx PP-OCRv5 mobile 输出维度 18385 (文件 18383 行 + 2)
+ * 反推得出, 适配大多数 PP-OCRv5 导出模型.
+ */
+@Getter
+public class Dictionary {
+
+    /** 字符表: idx 0 = CTC blank 占位, idx 1 = " " (半角空格), idx 2.. = 文件字符 */
+    private final List<String> characters;
+
+    private Dictionary(List<String> characters) {
+        this.characters = characters;
+    }
+
+    public int size() {
+        return characters.size();
+    }
+
+    public static Dictionary load(Path dictPath) throws IOException {
+        List<String> chars = new ArrayList<>();
+        // idx 0 = CTC blank 占位 (模型 output[0] = blank, 由 CtcDecoder 跳过)
+        chars.add("");
+        try (BufferedReader r = Files.newBufferedReader(dictPath, StandardCharsets.UTF_8)) {
+            String line;
+            while ((line = r.readLine()) != null) {
+                if (!line.isEmpty()) {
+                    chars.add(line);
+                }
+            }
+        }
+        return new Dictionary(chars);
+    }
+}
diff --git a/ry-api/ruoyi-business/src/main/java/com/ruoyi/business/ocr/core/ImageProcessor.java b/ry-api/ruoyi-business/src/main/java/com/ruoyi/business/ocr/core/ImageProcessor.java
new file mode 100644
index 0000000..efa23a8
--- /dev/null
+++ b/ry-api/ruoyi-business/src/main/java/com/ruoyi/business/ocr/core/ImageProcessor.java
@@ -0,0 +1,227 @@
+package com.ruoyi.business.ocr.core;
+
+import lombok.extern.slf4j.Slf4j;
+
+import javax.imageio.ImageIO;
+import java.awt.Graphics2D;
+import java.awt.RenderingHints;
+import java.awt.geom.AffineTransform;
+import java.awt.image.BufferedImage;
+import java.io.File;
+import java.io.IOException;
+import java.nio.file.Path;
+
+/**
+ * 图像预处理: 自动旋转 / 轻度增强 — 对齐 Python app.core.image_processor
+ * <p>
+ * 全部使用 Java 2D, 无需引入 OpenCV.
+ */
+@Slf4j
+public final class ImageProcessor {
+
+    private ImageProcessor() {}
+
+    /**
+     * 简易方向校正: 纵向图 (高 > 宽 × 1.2) 顺时针旋转 90°.
+     * <p>
+     * 复杂倾斜交给 ONNX 引擎自带的 textline orientation.
+     */
+    public static Path autoRotate(Path imgPath) {
+        Path p = imgPath;
+        BufferedImage img;
+        try {
+            img = ImageIO.read(p.toFile());
+        } catch (IOException e) {
+            log.warn("read image failed: {}", e.getMessage());
+            return p;
+        }
+        if (img == null) {
+            return p;
+        }
+        int h = img.getHeight();
+        int w = img.getWidth();
+        if (h > w * 1.2) {
+            BufferedImage rotated = rotate90Clockwise(img);
+            Path out = p.resolveSibling(stem(p) + "_rot.png");
+            try {
+                ImageIO.write(rotated, "png", out.toFile());
+                return out;
+            } catch (IOException e) {
+                log.warn("write rotated image failed: {}", e.getMessage());
+            }
+        }
+        return p;
+    }
+
+    /**
+     * 轻度增强: 灰度化 + (低对比度图) 自适应二值化.
+     */
+    public static Path enhance(Path imgPath) {
+        Path p = imgPath;
+        BufferedImage img;
+        try {
+            img = ImageIO.read(p.toFile());
+        } catch (IOException e) {
+            log.warn("read image failed: {}", e.getMessage());
+            return p;
+        }
+        if (img == null) {
+            return p;
+        }
+        BufferedImage gray = toGray(img);
+        double std = stddev(gray);
+        if (std < 50) {
+            BufferedImage binary = adaptiveThreshold(gray, 31, 10);
+            Path out = p.resolveSibling(stem(p) + "_enh.png");
+            try {
+                ImageIO.write(binary, "png", out.toFile());
+                return out;
+            } catch (IOException e) {
+                log.warn("write enhanced image failed: {}", e.getMessage());
+            }
+        }
+        return p;
+    }
+
+    // ---------- 内部 ----------
+
+    private static String stem(Path p) {
+        String name = p.getFileName().toString();
+        int dot = name.lastIndexOf('.');
+        return dot > 0 ? name.substring(0, dot) : name;
+    }
+
+    private static BufferedImage rotate90Clockwise(BufferedImage src) {
+        int w = src.getWidth();
+        int h = src.getHeight();
+        BufferedImage dst = new BufferedImage(h, w, src.getType() == 0 ? BufferedImage.TYPE_INT_RGB : src.getType());
+        Graphics2D g = dst.createGraphics();
+        AffineTransform tx = new AffineTransform();
+        tx.translate(h, 0);
+        tx.rotate(Math.toRadians(90));
+        g.setRenderingHint(RenderingHints.KEY_INTERPOLATION, RenderingHints.VALUE_INTERPOLATION_BILINEAR);
+        g.drawImage(src, tx, null);
+        g.dispose();
+        return dst;
+    }
+
+    private static BufferedImage toGray(BufferedImage src) {
+        if (src.getType() == BufferedImage.TYPE_BYTE_GRAY) return src;
+        BufferedImage gray = new BufferedImage(src.getWidth(), src.getHeight(), BufferedImage.TYPE_BYTE_GRAY);
+        Graphics2D g = gray.createGraphics();
+        g.drawImage(src, 0, 0, null);
+        g.dispose();
+        return gray;
+    }
+
+    /**
+     * 计算灰度图标准差 (判断是否低对比度)
+     * <p>
+     * 优化: 用 getRGB() 一次性读出整张图到 int[], 直接遍历 byte 提取亮度, 避免 Raster.getSample 逐元素调用 (慢 10x+).
+     */
+    private static double stddev(BufferedImage gray) {
+        int w = gray.getWidth();
+        int h = gray.getHeight();
+        int[] pixels = new int[w * h];
+        gray.getRGB(0, 0, w, h, pixels, 0, w);
+        long sum = 0;
+        long sumSq = 0;
+        long count = 0;
+        // 采样: 每 4 像素采样一次 (避免遍历几百万像素)
+        for (int i = 0; i < pixels.length; i += 16) {
+            int v = pixels[i] >>> 24 == 0 ? pixels[i] & 0xFF : (pixels[i] >> 16) & 0xFF; // 灰度图 R=G=B
+            // TYPE_BYTE_GRAY 的灰度值在 R/G/B 都一样, 取 R 即可
+            // 直接按 TYPE_BYTE_GRAY: ARGB 编码时 R 通道存的就是灰度
+            sum += v;
+            sumSq += (long) v * v;
+            count++;
+        }
+        if (count == 0) return 0;
+        double mean = (double) sum / count;
+        double variance = ((double) sumSq / count) - mean * mean;
+        return Math.sqrt(Math.max(0, variance));
+    }
+
+    /**
+     * 简易自适应二值化 (高斯加权 + 常数偏移).
+     * <p>
+     * 与 cv2.adaptiveThreshold(..., ADAPTIVE_THRESH_GAUSSIAN_C, ...) 行为近似.
+     * <p>
+     * 优化: 整图 getRGB 一次性读出, 用 byte[] 操作, 避免 Raster 逐元素访问.
+     */
+    private static BufferedImage adaptiveThreshold(BufferedImage gray, int blockSize, int C) {
+        int w = gray.getWidth();
+        int h = gray.getHeight();
+        int[] srcPixels = new int[w * h];
+        gray.getRGB(0, 0, w, h, srcPixels, 0, w);
+        // 提取灰度字节 (TYPE_BYTE_GRAY 在 BufferedImage 内部实际存为 TYPE_INT_ARGB 但 R 通道就是灰度)
+        byte[] srcGray = new byte[w * h];
+        for (int i = 0; i < srcPixels.length; i++) {
+            srcGray[i] = (byte) (srcPixels[i] & 0xFF);
+        }
+        // 先做 box blur (近似高斯)
+        byte[] blurredGray = boxBlurBytes(srcGray, w, h, blockSize);
+        BufferedImage out = new BufferedImage(w, h, BufferedImage.TYPE_BYTE_BINARY);
+        byte[] outData = new byte[w * h];
+        for (int y = 0; y < h; y++) {
+            int rowStart = y * w;
+            for (int x = 0; x < w; x++) {
+                int idx = rowStart + x;
+                int src = srcGray[idx] & 0xFF;
+                int bg = blurredGray[idx] & 0xFF;
+                int v = src - bg + 127 - C;
+                outData[idx] = (byte) (v > 127 ? 255 : 0);
+            }
+        }
+        out.getRaster().setDataElements(0, 0, w, h, outData);
+        return out;
+    }
+
+    /**
+     * 轻量 box blur (半径 = blockSize / 2) — 纯 byte[] 数组操作, 无 Raster 开销
+     */
+    private static byte[] boxBlurBytes(byte[] gray, int w, int h, int blockSize) {
+        int radius = Math.max(1, blockSize / 2);
+        // 构造积分图 (像素值累加, 0-255)
+        int[] integral = new int[w * h];
+        for (int y = 0; y < h; y++) {
+            int rowSum = 0;
+            int rowStart = y * w;
+            for (int x = 0; x < w; x++) {
+                rowSum += gray[rowStart + x] & 0xFF;
+                integral[rowStart + x] = rowSum + (y > 0 ? integral[rowStart + x - w] : 0);
+            }
+        }
+        byte[] out = new byte[w * h];
+        for (int y = 0; y < h; y++) {
+            int y1 = Math.max(0, y - radius);
+            int y2 = Math.min(h - 1, y + radius);
+            int rowStart = y * w;
+            for (int x = 0; x < w; x++) {
+                int x1 = Math.max(0, x - radius);
+                int x2 = Math.min(w - 1, x + radius);
+                int area = (x2 - x1 + 1) * (y2 - y1 + 1);
+                int sum = integral[y2 * w + x2];
+                if (x1 > 0) sum -= integral[y2 * w + (x1 - 1)];
+                if (y1 > 0) sum -= integral[(y1 - 1) * w + x2];
+                if (x1 > 0 && y1 > 0) sum += integral[(y1 - 1) * w + (x1 - 1)];
+                out[rowStart + x] = (byte) (sum / area);
+            }
+        }
+        return out;
+    }
+
+    /** 加载 BufferedImage (封装异常) */
+    public static BufferedImage read(Path p) throws IOException {
+        return ImageIO.read(p.toFile());
+    }
+
+    /** 保存 BufferedImage (封装异常) */
+    public static void write(BufferedImage img, Path p) throws IOException {
+        File f = p.toFile();
+        if (f.getParentFile() != null) {
+            f.getParentFile().mkdirs();
+        }
+        ImageIO.write(img, "png", f);
+    }
+}
diff --git a/ry-api/ruoyi-business/src/main/java/com/ruoyi/business/ocr/core/OcrEngine.java b/ry-api/ruoyi-business/src/main/java/com/ruoyi/business/ocr/core/OcrEngine.java
new file mode 100644
index 0000000..019b55b
--- /dev/null
+++ b/ry-api/ruoyi-business/src/main/java/com/ruoyi/business/ocr/core/OcrEngine.java
@@ -0,0 +1,231 @@
+package com.ruoyi.business.ocr.core;
+
+import ai.onnxruntime.OrtEnvironment;
+import ai.onnxruntime.OrtException;
+import com.ruoyi.business.ocr.config.OcrProperties;
+import com.ruoyi.business.ocr.exception.OcrTimeoutException;
+import com.ruoyi.business.ocr.OcrLine;
+import jakarta.annotation.PostConstruct;
+import jakarta.annotation.PreDestroy;
+import lombok.extern.slf4j.Slf4j;
+import org.springframework.stereotype.Component;
+
+import javax.imageio.ImageIO;
+import java.awt.image.BufferedImage;
+import java.io.IOException;
+import java.nio.file.Files;
+import java.nio.file.Path;
+import java.util.ArrayList;
+import java.util.Collections;
+import java.util.List;
+import java.util.concurrent.*;
+
+/**
+ * OCR 引擎单例 — 对齐 Python app.core.ocr_engine
+ * <p>
+ * 单例 + 单页超时 (Future.get(timeout)).
+ * 底层: PaddleOCR ONNX Runtime (det + rec).
+ */
+@Slf4j
+@Component
+public class OcrEngine {
+
+    private final OcrProperties props;
+    private final ExecutorService executor = Executors.newFixedThreadPool(2, r -> {
+        Thread t = new Thread(r, "ocr-worker");
+        t.setDaemon(true);
+        return t;
+    });
+
+    private OrtEnvironment ortEnv;
+    private TextDetector detector;
+    private TextRecognizer recognizer;
+
+    private volatile boolean ready = false;
+
+    public OcrEngine(OcrProperties props) {
+        this.props = props;
+    }
+
+    /**
+     * 启动预热: 加载 ONNX Session
+     */
+    @PostConstruct
+    public synchronized void warmup() {
+        try {
+            Path modelsDir = resolveModelsDir();
+            Path detPath = modelsDir.resolve("det.onnx");
+            Path recPath = modelsDir.resolve("rec.onnx");
+            Path dictPath = modelsDir.resolve("ppocr_keys_v1.txt");
+
+            if (!Files.exists(detPath) || !Files.exists(recPath) || !Files.exists(dictPath)) {
+                log.warn("OCR 模型未找到 ({}/det.onnx + rec.onnx + ppocr_keys_v1.txt), 引擎未就绪, /health 返回 degraded", modelsDir);
+                return;
+            }
+
+            log.info("加载 OCR 模型 from {} (version={})", modelsDir, props.getOcr().getModelVersion());
+            long t0 = System.currentTimeMillis();
+            this.ortEnv = OrtEnvironment.getEnvironment();
+            this.detector = new TextDetector(ortEnv, detPath, resolveDetMaxSide());
+            Dictionary dict = Dictionary.load(dictPath);
+            this.recognizer = new TextRecognizer(ortEnv, recPath, dict, resolveRecHeight(), resolveRecMaxW());
+            this.ready = true;
+            log.info("OCR 引擎就绪, 耗时 {}ms (recHeight={}, recMaxW={}, detMaxSide={})",
+                    System.currentTimeMillis() - t0, resolveRecHeight(), resolveRecMaxW(), resolveDetMaxSide());
+        } catch (Exception e) {
+            log.warn("OCR 引擎初始化失败: {}", e.getMessage(), e);
+        }
+    }
+
+    /**
+     * 解析模型目录: 相对路径 → classpath:models/, 绝对路径 → 直读
+     */
+    private Path resolveModelsDir() {
+        String dir = props.getOcr().getModelsDir();
+        Path p = Path.of(dir);
+        if (p.isAbsolute()) return p;
+
+        // 部署约定: 模型与 jar 同目录 (user.dir = java -jar 启动目录, 见 ry.sh 的 `pwd`)
+        Path workDir = Path.of(System.getProperty("user.dir"), dir);
+        if (hasModels(workDir)) return workDir;
+
+        // classpath: resources/models/ (开发环境备选)
+        try {
+            java.net.URL url = getClass().getClassLoader().getResource(dir);
+            if (url != null && "file".equals(url.getProtocol())) {
+                return Path.of(url.toURI());
+            }
+        } catch (Exception ignored) {}
+
+        // IDE 源码目录回退
+        return Path.of("src/main/resources", dir);
+    }
+
+    /** det.onnx + rec.onnx + ppocr_keys_v1.txt 齐全才算有效模型目录 */
+    private boolean hasModels(Path modelsDir) {
+        return Files.exists(modelsDir.resolve("det.onnx"))
+                && Files.exists(modelsDir.resolve("rec.onnx"))
+                && Files.exists(modelsDir.resolve("ppocr_keys_v1.txt"));
+    }
+
+    /** 解析 CRNN 输入高度: 显式 rec-height > modelVersion 默认 */
+    private int resolveRecHeight() {
+        Integer override = props.getOcr().getRecHeight();
+        if (override != null) return override;
+        return switch (props.getOcr().getModelVersion()) {
+            case "v5_server", "v5_mobile" -> 48;
+            case "v4_mobile" -> 32;
+            default -> 48;
+        };
+    }
+
+    private int resolveRecMaxW() {
+        Integer override = props.getOcr().getRecMaxWidth();
+        if (override != null) return override;
+        return 320;
+    }
+
+    private int resolveDetMaxSide() {
+        Integer override = props.getOcr().getDetMaxSide();
+        if (override != null) return override;
+        return switch (props.getOcr().getModelVersion()) {
+            case "v5_server", "v5_mobile" -> 800; // 优化: v5 原生 960 → 800, 算力 -31%
+            case "v4_mobile" -> 960;
+            default -> 960;
+        };
+    }
+
+    @PreDestroy
+    public void close() {
+        try {
+            if (recognizer != null) recognizer.close();
+            if (detector != null) detector.close();
+        } catch (Exception e) {
+            log.warn("close engine error: {}", e.getMessage());
+        }
+        executor.shutdownNow();
+    }
+
+    public boolean isReady() {
+        return ready;
+    }
+
+    /**
+     * 识别单张图片 — 带单页超时
+     */
+    public List<OcrLine> recognize(Path imagePath) {
+        if (!ready) {
+            throw new IllegalStateException("OCR 引擎未就绪, 请检查 models/ 目录下是否有 det.onnx / rec.onnx / ppocr_keys_v1.txt");
+        }
+        int timeoutSec = props.getOcr().getPageTimeoutS();
+        Future<List<OcrLine>> future = executor.submit(() -> doRecognize(imagePath));
+        try {
+            return future.get(timeoutSec, TimeUnit.SECONDS);
+        } catch (TimeoutException e) {
+            future.cancel(true);
+            throw new OcrTimeoutException("OCR 识别超时 (" + timeoutSec + "秒): " + imagePath);
+        } catch (InterruptedException e) {
+            Thread.currentThread().interrupt();
+            throw new OcrRuntimeException("OCR 中断: " + e.getMessage());
+        } catch (ExecutionException e) {
+            Throwable cause = e.getCause();
+            if (cause instanceof OcrTimeoutException) throw (OcrTimeoutException) cause;
+            if (cause instanceof OcrRuntimeException) throw (OcrRuntimeException) cause;
+            throw new OcrRuntimeException("OCR 执行异常: " + (cause == null ? e.getMessage() : cause.getMessage()), cause);
+        }
+    }
+
+    private List<OcrLine> doRecognize(Path imagePath) {
+        try {
+            BufferedImage img = ImageIO.read(imagePath.toFile());
+            if (img == null) return Collections.emptyList();
+
+            // 1. 检测
+            List<List<Float>> boxes = detector.detect(img);
+            // 2. 过滤 + 收集 crops (排除 height<8 或 width<5 的噪点 — 不会影响字段抽取)
+            List<BufferedImage> crops = new ArrayList<>();
+            List<List<Float>> validBoxes = new ArrayList<>();
+            for (List<Float> box : boxes) {
+                int[] bb = DbPostProcessor.bbox(box);
+                int x1 = Math.max(0, bb[0]);
+                int y1 = Math.max(0, bb[1]);
+                int x2 = Math.min(img.getWidth(), bb[2]);
+                int y2 = Math.min(img.getHeight(), bb[3]);
+                int w = x2 - x1, h = y2 - y1;
+                if (w < 5 || h < 8) continue;  // 过小 — 噪点
+                if (w < 3 || h < 3) continue;   // 原安全检查
+                crops.add(img.getSubimage(x1, y1, w, h));
+                validBoxes.add(box);
+            }
+
+            // 3. 批量推理 (性能关键: ONNX 一次推理处理所有 crop)
+            List<TextRecognizer.RecognizedText> rts = recognizer.recognizeBatch(crops);
+
+            // 4. 配对 boxes + texts
+            List<OcrLine> lines = new ArrayList<>();
+            for (int i = 0; i < validBoxes.size(); i++) {
+                TextRecognizer.RecognizedText rt = rts.get(i);
+                if (rt.text() != null && !rt.text().isEmpty()) {
+                    List<Float> poly = validBoxes.get(i);
+                    List<List<Float>> boxList = new ArrayList<>();
+                    for (int k = 0; k < poly.size(); k += 2) {
+                        List<Float> p = new ArrayList<>();
+                        p.add(poly.get(k));
+                        p.add(poly.get(k + 1));
+                        boxList.add(p);
+                    }
+                    lines.add(new OcrLine(rt.text().trim(), rt.confidence(), boxList));
+                }
+            }
+            return lines;
+        } catch (OrtException | IOException e) {
+            throw new OcrRuntimeException("OCR 执行异常: " + e.getMessage(), e);
+        }
+    }
+
+    /** 内部异常, 避免暴露 ONNX 细节 */
+    public static class OcrRuntimeException extends RuntimeException {
+        public OcrRuntimeException(String message) { super(message); }
+        public OcrRuntimeException(String message, Throwable cause) { super(message, cause); }
+    }
+}
diff --git a/ry-api/ruoyi-business/src/main/java/com/ruoyi/business/ocr/core/PdfProcessor.java b/ry-api/ruoyi-business/src/main/java/com/ruoyi/business/ocr/core/PdfProcessor.java
new file mode 100644
index 0000000..ab42988
--- /dev/null
+++ b/ry-api/ruoyi-business/src/main/java/com/ruoyi/business/ocr/core/PdfProcessor.java
@@ -0,0 +1,86 @@
+package com.ruoyi.business.ocr.core;
+
+import com.ruoyi.business.ocr.config.OcrProperties;
+import lombok.RequiredArgsConstructor;
+import lombok.extern.slf4j.Slf4j;
+import org.apache.pdfbox.pdmodel.PDDocument;
+import org.apache.pdfbox.rendering.PDFRenderer;
+import org.apache.pdfbox.text.PDFTextStripper;
+import org.springframework.stereotype.Component;
+
+import java.awt.image.BufferedImage;
+import java.io.IOException;
+import java.nio.file.Path;
+import java.util.ArrayList;
+import java.util.List;
+
+/**
+ * PDF → 图片 — 对齐 Python app.core.pdf_processor
+ * <p>
+ * 使用 PDFBox (无需 poppler 等系统依赖).
+ */
+@Slf4j
+@Component
+@RequiredArgsConstructor
+public class PdfProcessor {
+
+    private final OcrProperties props;
+
+    /**
+     * 把 PDF 每页渲染成 PNG, 返回临时文件路径列表.
+     * <p>
+     * 输出目录: {pdf.parent}/.{pdf.stem}_pages/page_{idx:03d}.png
+     */
+    public List<Path> pdfToImages(Path pdfPath) throws IOException {
+        int dpi = props.getUpload().getPdfDpi();
+        Path outDir = pdfPath.getParent().resolve("." + stem(pdfPath) + "_pages");
+        outDir.toFile().mkdirs();
+
+        List<Path> saved = new ArrayList<>();
+        try (PDDocument doc = PDDocument.load(pdfPath.toFile())) {
+            PDFRenderer renderer = new PDFRenderer(doc);
+            renderer.setSubsamplingAllowed(true);  // 大图降采样, 提速 + 省内存
+            int pageCount = doc.getNumberOfPages();
+            for (int idx = 0; idx < pageCount; idx++) {
+                BufferedImage img = renderer.renderImageWithDPI(idx, dpi);
+                Path outPath = outDir.resolve(String.format("page_%03d.png", idx + 1));
+                ImageProcessor.write(img, outPath);
+                saved.add(outPath);
+            }
+        }
+        log.info("PDF 转图片: {} → {} 页 (dpi={})", pdfPath.getFileName(), saved.size(), dpi);
+        return saved;
+    }
+
+    private static String stem(Path p) {
+        String name = p.getFileName().toString();
+        int dot = name.lastIndexOf('.');
+        return dot > 0 ? name.substring(0, dot) : name;
+    }
+
+    /**
+     * 抽取 PDF 内嵌文本（pdftotext 等价）— 用于"不调用 ONNX"的 fast path
+     * <p>
+     * 适用: 电子发票 / 数电票 PDF (含真实可复制文本层).
+     * 扫描件 PDF 抽出来为空或字符极少, 调用方应回退到 OCR 流程.
+     *
+     * @return 抽取到的纯文本 (trim 后); 若文件无文本层, 返回空字符串
+     */
+    public String extractText(Path pdfPath) throws IOException {
+        try (PDDocument doc = PDDocument.load(pdfPath.toFile())) {
+            PDFTextStripper stripper = new PDFTextStripper();
+            // 逐页拼接
+            StringBuilder sb = new StringBuilder();
+            int pageCount = doc.getNumberOfPages();
+            for (int i = 1; i <= pageCount; i++) {
+                stripper.setStartPage(i);
+                stripper.setEndPage(i);
+                sb.append(stripper.getText(doc));
+                if (i < pageCount) sb.append('\n');
+            }
+            String text = sb.toString().trim();
+            log.info("PDF 内嵌文本抽取: {} → {} 字符 ({} 页)", pdfPath.getFileName(), text.length(), pageCount);
+            return text;
+        }
+    }
+}
diff --git a/ry-api/ruoyi-business/src/main/java/com/ruoyi/business/ocr/core/TextDetector.java b/ry-api/ruoyi-business/src/main/java/com/ruoyi/business/ocr/core/TextDetector.java
new file mode 100644
index 0000000..9a220e9
--- /dev/null
+++ b/ry-api/ruoyi-business/src/main/java/com/ruoyi/business/ocr/core/TextDetector.java
@@ -0,0 +1,156 @@
+package com.ruoyi.business.ocr.core;
+
+import ai.onnxruntime.OrtEnvironment;
+import ai.onnxruntime.OrtException;
+import ai.onnxruntime.OrtSession;
+import ai.onnxruntime.OrtSession.SessionOptions;
+import ai.onnxruntime.OnnxTensor;
+import ai.onnxruntime.OrtSession.SessionOptions.ExecutionMode;
+import ai.onnxruntime.OrtSession.SessionOptions.OptLevel;
+import lombok.extern.slf4j.Slf4j;
+
+import java.awt.image.BufferedImage;
+import java.nio.FloatBuffer;
+import java.nio.file.Path;
+import java.util.ArrayList;
+import java.util.Collections;
+import java.util.List;
+import java.util.Map;
+
+/**
+ * PaddleOCR 检测 (DB 算法) — ONNX Runtime 推理.
+ * <p>
+ * 输入: [1, 3, H, W] 归一化图 (mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])
+ * 输出: 概率图 [1, 1, H, W]
+ */
+@Slf4j
+public class TextDetector implements AutoCloseable {
+
+    private final OrtEnvironment env;
+    private final OrtSession session;
+    private final float[] mean = {0.485f, 0.456f, 0.406f};
+    private final float[] std = {0.229f, 0.224f, 0.225f};
+    /** DB 二值化阈值 */
+    private final float dbThresh = 0.3f;
+    /** DB box 阈值 */
+    private final float boxThresh = 0.5f;
+    /** unclip 膨胀系数 */
+    private final float unclipRatio = 1.6f;
+    /** 最长边限制 — 通过构造参数传入 (v5 默认 800, v4 默认 960) */
+    private final int maxSideLen;
+
+    public TextDetector(OrtEnvironment env, Path modelPath, int maxSideLen) throws OrtException {
+        this.maxSideLen = maxSideLen;
+        this.env = env;
+        SessionOptions opts = new SessionOptions();
+        opts.setExecutionMode(ExecutionMode.PARALLEL);
+        opts.setOptimizationLevel(OptLevel.ALL_OPT);
+        int threads = Math.max(2, Runtime.getRuntime().availableProcessors());
+        opts.setIntraOpNumThreads(threads);
+        opts.setInterOpNumThreads(threads);
+        this.session = env.createSession(modelPath.toString(), opts);
+        log.info("DB 检测器加载完成: {}, threads={}", modelPath.getFileName(), threads);
+    }
+
+    /**
+     * 检测: 返回多边形 (每个 polygon 是 4-8 个 [x, y] 点)
+     */
+    public List<List<Float>> detect(BufferedImage img) throws OrtException {
+        // 1. 预处理 (resize, pad, normalize)
+        int origH = img.getHeight();
+        int origW = img.getWidth();
+        // 按最长边缩放
+        int targetH = origH, targetW = origW;
+        int maxSide = Math.max(origH, origW);
+        if (maxSide > maxSideLen) {
+            float ratio = (float) maxSideLen / maxSide;
+            targetH = Math.round(origH * ratio);
+            targetW = Math.round(origW * ratio);
+        }
+        // pad 到 32 倍数
+        int padH = (32 - targetH % 32) % 32;
+        int padW = (32 - targetW % 32) % 32;
+        int inputH = targetH + padH;
+        int inputW = targetW + padW;
+
+        BufferedImage resized = resize(img, targetW, targetH);
+        float[] inputData = new float[3 * inputH * inputW];
+        // CHW 归一化 + pad
+        int[] pixels = new int[targetW * targetH];
+        resized.getRGB(0, 0, targetW, targetH, pixels, 0, targetW);
+        // CHW
+        for (int c = 0; c < 3; c++) {
+            for (int y = 0; y < targetH; y++) {
+                for (int x = 0; x < targetW; x++) {
+                    int argb = pixels[y * targetW + x];
+                    int v;
+                    switch (c) {
+                        case 0: v = (argb >> 16) & 0xFF; break;
+                        case 1: v = (argb >> 8) & 0xFF; break;
+                        default: v = argb & 0xFF;
+                    }
+                    int idx = c * inputH * inputW + y * inputW + x;
+                    inputData[idx] = ((float) v / 255f - mean[c]) / std[c];
+                }
+            }
+            // pad 行: 已经是 0, 因为 FloatBuffer 默认 0
+        }
+
+        // 2. 推理
+        long[] shape = {1, 3, inputH, inputW};
+        OnnxTensor inputTensor = OnnxTensor.createTensor(env, FloatBuffer.wrap(inputData), shape);
+        Map<String, OnnxTensor> inputs = Collections.singletonMap("x", inputTensor);
+
+        List<List<Float>> boxes;
+        try (OrtSession.Result result = session.run(inputs)) {
+            OnnxTensor outTensor = (OnnxTensor) result.get(0);
+            float[][] prob = extractProbMap(outTensor, inputH, inputW);
+            // 3. DB 后处理
+            boxes = DbPostProcessor.postProcess(prob, targetH, targetW, origH, origW,
+                    dbThresh, boxThresh, unclipRatio);
+        } finally {
+            inputTensor.close();
+        }
+        return boxes;
+    }
+
+    /**
+     * 从 ONNX tensor 抽取概率图为 [H, W] float[][]
+     * <p>
+     * PaddleOCR 检测模型输出可能是 float[1][1][H][W] / float[1][H][W] / FloatBuffer 等.
+     */
+    private static float[][] extractProbMap(OnnxTensor tensor, int h, int w) throws OrtException {
+        Object val = tensor.getValue();
+        if (val instanceof float[][][][]) {
+            return ((float[][][][]) val)[0][0];
+        } else if (val instanceof float[][][]) {
+            return ((float[][][]) val)[0];
+        } else if (val instanceof float[][]) {
+            return (float[][]) val;
+        } else if (val instanceof FloatBuffer) {
+            FloatBuffer buf = (FloatBuffer) val;
+            float[][] map = new float[h][w];
+            for (int y = 0; y < h; y++) {
+                for (int x = 0; x < w; x++) {
+                    map[y][x] = buf.get(y * w + x);
+                }
+            }
+            return map;
+        }
+        throw new IllegalStateException("不支持的 ONNX 输出类型: " + (val == null ? "null" : val.getClass()));
+    }
+
+    private static BufferedImage resize(BufferedImage src, int w, int h) {
+        java.awt.Image tmp = src.getScaledInstance(w, h, java.awt.Image.SCALE_SMOOTH);
+        BufferedImage dst = new BufferedImage(w, h, BufferedImage.TYPE_INT_RGB);
+        java.awt.Graphics2D g = dst.createGraphics();
+        g.drawImage(tmp, 0, 0, null);
+        g.dispose();
+        return dst;
+    }
+
+    @Override
+    public void close() throws OrtException {
+        session.close();
+    }
+}
diff --git a/ry-api/ruoyi-business/src/main/java/com/ruoyi/business/ocr/core/TextRecognizer.java b/ry-api/ruoyi-business/src/main/java/com/ruoyi/business/ocr/core/TextRecognizer.java
new file mode 100644
index 0000000..6ce134f
--- /dev/null
+++ b/ry-api/ruoyi-business/src/main/java/com/ruoyi/business/ocr/core/TextRecognizer.java
@@ -0,0 +1,181 @@
+package com.ruoyi.business.ocr.core;
+
+import ai.onnxruntime.OnnxTensor;
+import ai.onnxruntime.OrtEnvironment;
+import ai.onnxruntime.OrtException;
+import ai.onnxruntime.OrtSession;
+import ai.onnxruntime.OrtSession.SessionOptions;
+import ai.onnxruntime.OrtSession.SessionOptions.ExecutionMode;
+import ai.onnxruntime.OrtSession.SessionOptions.OptLevel;
+import lombok.extern.slf4j.Slf4j;
+
+import java.awt.image.BufferedImage;
+import java.nio.FloatBuffer;
+import java.nio.file.Path;
+import java.util.ArrayList;
+import java.util.Collections;
+import java.util.List;
+import java.util.Map;
+
+/**
+ * PaddleOCR 识别 (CRNN + CTC) — ONNX Runtime 推理.
+ * <p>
+ * 输入: [1, 3, H, W] 归一化图 (mean=0.5, std=0.5)
+ * 输出: [1, T, N] (T=序列长度, N=字典大小+1 含 blank)
+ * <p>
+ * 当前使用 PP-OCRv4 mobile: H=32, maxW=320 (v4 标准, v5 是 H=48).
+ */
+@Slf4j
+public class TextRecognizer implements AutoCloseable {
+
+    private final OrtEnvironment env;
+    private final OrtSession session;
+    private final Dictionary dict;
+    private final float[] mean = {0.5f, 0.5f, 0.5f};
+    private final float[] std = {0.5f, 0.5f, 0.5f};
+    /** 输入最大宽 — 通过构造参数传入 */
+    private final int maxW;
+    /** 输入高度 — 通过构造参数传入 (v5=48, v4=32) */
+    private final int targetH;
+
+    public TextRecognizer(OrtEnvironment env, Path modelPath, Dictionary dict, int targetH, int maxW) throws OrtException {
+        this.env = env;
+        this.dict = dict;
+        this.targetH = targetH;
+        this.maxW = maxW;
+        SessionOptions opts = new SessionOptions();
+        // 并行执行模式 + 全图优化 (节点融合/常量折叠) — CRNN 是主要瓶颈, 收益最大
+        opts.setExecutionMode(ExecutionMode.PARALLEL);
+        opts.setOptimizationLevel(OptLevel.ALL_OPT);
+        // CRNN 单图推理 op 数少, intra-op 多核并行收益高; inter-op 多图并发 (无 batch 时无效)
+        int threads = Math.max(2, Runtime.getRuntime().availableProcessors());
+        opts.setIntraOpNumThreads(threads);
+        opts.setInterOpNumThreads(threads);
+        this.session = env.createSession(modelPath.toString(), opts);
+        log.info("CRNN 识别器加载完成: {}, 字典={} 字符, threads={}", modelPath.getFileName(), dict.size(), threads);
+    }
+
+    /**
+     * 识别一张裁剪图, 返回 (文本, 置信度)
+     */
+    public RecognizedText recognize(BufferedImage crop) throws OrtException {
+        // 复用 batch 推理, 单图 = batch=1
+        List<RecognizedText> results = recognizeBatch(Collections.singletonList(crop));
+        return results.get(0);
+    }
+
+    /**
+     * 批量识别多张裁剪图 — 性能关键.
+     * <p>
+     * 核心优化: 把所有 crop pad 到 batch 内统一宽度 (maxW), 一次性喂给 ONNX.
+     * ONNX 内部用 8 线程并行算所有样本, 单图推理省掉 33 次 kernel launch.
+     * <p>
+     * 输入: 每张 crop 任意宽度 → resize 高 48 + 宽按比例 ≤ maxW
+     * 输出: 与输入 crops 一一对应的 RecognizedText
+     */
+    public List<RecognizedText> recognizeBatch(List<BufferedImage> crops) throws OrtException {
+        int B = crops.size();
+        if (B == 0) return Collections.emptyList();
+
+        // targetH 来自构造参数 (v5=48, v4=32, 通过 OcrEngine 从配置读取)
+
+        // 1. 计算每张 crop 的目标宽度 + batch 内最大宽度
+        int[] targetWs = new int[B];
+        int batchMaxW = 0;
+        for (int i = 0; i < B; i++) {
+            BufferedImage c = crops.get(i);
+            float ratio = (float) targetH / c.getHeight();
+            int w = Math.min(maxW, Math.max(1, Math.round(c.getWidth() * ratio)));
+            // pad 到 8 倍数 (CRNN 下采样 8x)
+            w = ((w + 7) / 8) * 8;
+            targetWs[i] = w;
+            if (w > batchMaxW) batchMaxW = w;
+        }
+
+        // 2. 拼成 [B, 3, 48, batchMaxW] — 每张图只在 [0..targetW] 区间写, 其余为 0
+        float[] inputData = new float[B * 3 * targetH * batchMaxW];
+        for (int b = 0; b < B; b++) {
+            int w = targetWs[b];
+            if (w == 0) continue;
+            BufferedImage resized = resize(crops.get(b), w, targetH);
+            int[] pixels = new int[w * targetH];
+            resized.getRGB(0, 0, w, targetH, pixels, 0, w);
+            int bOffset = b * 3 * targetH * batchMaxW;
+            for (int c = 0; c < 3; c++) {
+                int cOffset = bOffset + c * targetH * batchMaxW;
+                int meanC = (int) (mean[c] * 255);
+                int stdC = (int) (std[c] * 255);
+                for (int y = 0; y < targetH; y++) {
+                    int rowStart = cOffset + y * batchMaxW;
+                    int pixRowStart = y * w;
+                    for (int x = 0; x < w; x++) {
+                        int argb = pixels[pixRowStart + x];
+                        int v;
+                        switch (c) {
+                            case 0: v = (argb >> 16) & 0xFF; break;
+                            case 1: v = (argb >> 8) & 0xFF; break;
+                            default: v = argb & 0xFF;
+                        }
+                        // (v/255 - mean) / std, 避免浮点除法 (mean/std 都是 0.5)
+                        inputData[rowStart + x] = (v - meanC) / (stdC * 1.0f);
+                    }
+                }
+            }
+        }
+
+        // 3. 推理
+        long[] shape = {B, 3, targetH, batchMaxW};
+        OnnxTensor inputTensor = OnnxTensor.createTensor(env, FloatBuffer.wrap(inputData), shape);
+        Map<String, OnnxTensor> inputs = Collections.singletonMap("x", inputTensor);
+
+        List<RecognizedText> results = new ArrayList<>(B);
+        try (OrtSession.Result result = session.run(inputs)) {
+            OnnxTensor outTensor = (OnnxTensor) result.get(0);
+            Object val = outTensor.getValue();
+            float[][][] logits3d;
+            if (val instanceof float[][][]) {
+                logits3d = (float[][][]) val;
+            } else if (val instanceof float[][][][]) {
+                logits3d = ((float[][][][]) val)[0];
+            } else {
+                throw new IllegalStateException("不支持的 ONNX 输出类型: " + (val == null ? "null" : val.getClass()));
+            }
+            for (int b = 0; b < B; b++) {
+                results.add(CtcDecoder.decode(logits3d[b], dict));
+            }
+        } finally {
+            inputTensor.close();
+        }
+        return results;
+    }
+
+    /**
+     * 从 ONNX tensor 抽取 logits 为 [T, N] float[][]
+     */
+    private static float[][] extractLogits(OnnxTensor tensor) throws OrtException {
+        Object val = tensor.getValue();
+        if (val instanceof float[][][]) {
+            return ((float[][][]) val)[0];
+        } else if (val instanceof float[][]) {
+            return (float[][]) val;
+        }
+        throw new IllegalStateException("不支持的 ONNX 输出类型: " + (val == null ? "null" : val.getClass()));
+    }
+
+    private static BufferedImage resize(BufferedImage src, int w, int h) {
+        java.awt.Image tmp = src.getScaledInstance(w, h, java.awt.Image.SCALE_SMOOTH);
+        BufferedImage dst = new BufferedImage(w, h, BufferedImage.TYPE_INT_RGB);
+        java.awt.Graphics2D g = dst.createGraphics();
+        g.drawImage(tmp, 0, 0, null);
+        g.dispose();
+        return dst;
+    }
+
+    @Override
+    public void close() throws OrtException {
+        session.close();
+    }
+
+    /** 识别结果 */
+    public record RecognizedText(String text, double confidence) {}
+}
diff --git a/ry-api/ruoyi-business/src/main/java/com/ruoyi/business/ocr/exception/OcrTimeoutException.java b/ry-api/ruoyi-business/src/main/java/com/ruoyi/business/ocr/exception/OcrTimeoutException.java
new file mode 100644
index 0000000..37d27c9
--- /dev/null
+++ b/ry-api/ruoyi-business/src/main/java/com/ruoyi/business/ocr/exception/OcrTimeoutException.java
@@ -0,0 +1,17 @@
+package com.ruoyi.business.ocr.exception;
+
+/**
+ * OCR 识别超时异常 — 对齐 Python app.core.ocr_engine.OCRTimeout
+ * <p>
+ * 单页 OCR 超时 / 整流程超时时抛出, 由 RecognizeService 转成 error_code="timeout".
+ */
+public class OcrTimeoutException extends RuntimeException {
+
+    public OcrTimeoutException(String message) {
+        super(message);
+    }
+
+    public OcrTimeoutException(String message, Throwable cause) {
+        super(message, cause);
+    }
+}
diff --git a/ry-api/ruoyi-business/src/main/java/com/ruoyi/business/ocr/service/InvoiceExtractor.java b/ry-api/ruoyi-business/src/main/java/com/ruoyi/business/ocr/service/InvoiceExtractor.java
new file mode 100644
index 0000000..470d668
--- /dev/null
+++ b/ry-api/ruoyi-business/src/main/java/com/ruoyi/business/ocr/service/InvoiceExtractor.java
@@ -0,0 +1,434 @@
+package com.ruoyi.business.ocr.service;
+
+import com.ruoyi.business.ocr.InvoiceFields;
+import com.ruoyi.business.ocr.OcrLine;
+import com.ruoyi.business.ocr.util.AmountUtils;
+import lombok.extern.slf4j.Slf4j;
+import org.springframework.stereotype.Service;
+
+import java.util.*;
+import java.util.regex.Matcher;
+import java.util.regex.Pattern;
+
+/**
+ * 从 OCR 文本/行里抽取发票字段 — 对齐 Python app.services.invoice_extractor
+ * <p>
+ * 适配中国大陆 增值税发票（电子普票 / 专票 / 电子专票 / 数电票）.
+ * <p>
+ * 关键策略:
+ *   - 主体按 OCR box 坐标判断归属 (左右两栏)
+ *   - 名称提取加 stop word, 避免单行文本混淆
+ *   - 金额兜底: tax + pretax = total 组合搜索
+ */
+@Slf4j
+@Service
+public class InvoiceExtractor {
+
+    // ---------- 发票类型 (按长度降序, 优先匹配最长前缀, 避免 "电子发票（增值税专用发票）" 被 "增值税专用发票" 抢先命中) ----------
+    private static final List<String> INVOICE_TYPES = Arrays.asList(
+            "电子发票（增值税专用发票）",  // 13
+            "电子发票（增值税普通发票）",  // 13
+            "增值税电子专用发票",         // 9
+            "增值税电子普通发票",         // 9
+            "增值税专用发票",             // 7
+            "增值税普通发票",             // 7
+            "通用机打发票",               // 6
+            "数电票（电子发票）",         // 8
+            "数电票",                     // 3
+            "电子发票"                    // 4
+    );
+
+    // ---------- 发票号码 ----------
+    private static final Pattern NO_PATTERN = Pattern.compile(
+            "(?:发\\s*票\\s*号\\s*码|号\\s*码|No\\.?|号)\\s*[:：]?\\s*(\\d{8,20})",
+            Pattern.CASE_INSENSITIVE
+    );
+
+    // ---------- 发票代码 ----------
+    private static final Pattern CODE_PATTERN = Pattern.compile(
+            "(?:发\\s*票\\s*代\\s*码|代\\s*码)\\s*[:：]?\\s*(\\d{10,12}|\\d{8,12})"
+    );
+
+    // ---------- 开票日期 ----------
+    private static final Pattern DATE_PATTERN = Pattern.compile(
+            "(?:开\\s*票\\s*日\\s*期|日\\s*期)\\s*[:：]?\\s*" +
+                    "(\\d{4})\\s*[年/\\.]\\s*(\\d{1,2})\\s*[月/\\.]\\s*(\\d{1,2})"
+    );
+
+    // ---------- 纳税人识别号 (至少 1 个字母, 排除纯数字发票号) ----------
+    private static final Pattern TAX_NO_PATTERN = Pattern.compile("((?=[0-9A-Z]*[A-Z])[0-9A-Z]{18})");
+
+    // ---------- 主体标签 ----------
+    private static final Pattern BUYER_LABEL = Pattern.compile("购\\s*买\\s*方\\s*(?:信\\s*息|名\\s*称|)");
+    private static final Pattern SELLER_LABEL = Pattern.compile("销\\s*售\\s*方\\s*(?:信\\s*息|名\\s*称|)");
+
+    // ---------- 名称 (带 stop word 截断) ----------
+    private static final String NAME_STOP = "(?:销售方|购买方|统一社会信用|纳税人|项目名称|规格型号|^单位$|^数量$|^单价$|^金额|^税率|^税额|备注|收款人|复核|开票人|价税合计|小写|大写)";
+    private static final Pattern NAME_PATTERN = Pattern.compile(
+            "名\\s*称\\s*[:：]\\s*" +
+                    "((?:(?!" + NAME_STOP + ")[^\\n\\r]){2,60}?(?:公司|商店|厂|店|部|中心|工作室))"
+    );
+    // 不依赖 "名称：" 前缀 — 用于 PDF 内嵌文本拆字版式 (label 和 value 分两段)
+    // 不强制 lookback 拒绝中文 (OCR 行可能整行连在一起如 "名称北京国钜...公司"), 靠 cleanName 截 stop word 过滤杂质
+    private static final Pattern COMPANY_PATTERN = Pattern.compile(
+            "([一-龥A-Za-z0-9（）()·\\-]{2,30}(?:公司|商店|厂|店|部|中心|工作室))"
+    );
+
+    // ---------- 数字候选 ----------
+    private static final Pattern DECIMAL_PATTERN = Pattern.compile("(\\d+\\.\\d{2})");
+
+    /**
+     * 入口: 从文本 + (可选) 行列表抽取
+     *
+     * @param text        OCR 全文
+     * @param lines       OCR 行列表 (含 box)
+     * @param qrTotalAmount  QR 解出的金额 (权威). 不为 null 时, 强制作为 total 用于 fallback 组合搜索.
+     */
+    public InvoiceFields extract(String text, List<OcrLine> lines, Double qrTotalAmount) {
+        String norm = norm(text);
+        InvoiceFields fields = new InvoiceFields();
+
+        fields.setInvoiceType(detectInvoiceType(norm));
+        fields.setInvoiceCode(extractInvoiceCode(norm));
+        fields.setInvoiceNo(extractInvoiceNo(norm));
+        fields.setInvoiceDate(extractDate(norm));
+
+        // 金额 — 优先用 QR 提供的 total (权威, 数电票/电子发票的价税合计在 QR 里),
+        //       QR 缺失时回退 OCR 的 TOTAL_PATTERN / 第一个数字
+        Double total;
+        if (qrTotalAmount != null) {
+            total = qrTotalAmount;
+        } else {
+            total = AmountUtils.extractTotalAmount(norm);
+        }
+        Double tax = AmountUtils.extractTaxAmount(norm);
+        Double pretax = AmountUtils.extractPretaxAmount(norm);
+
+        // 兜底 1: tax + pretax = total 组合搜索
+        if ((tax == null || pretax == null) && total != null) {
+            List<Double> candidates = new ArrayList<>();
+            Set<String> seen = new HashSet<>();
+            Matcher m = DECIMAL_PATTERN.matcher(norm);
+            while (m.find()) {
+                String s = m.group(1);
+                double v = Double.parseDouble(s);
+                if (v < total && seen.add(s)) {
+                    candidates.add(v);
+                }
+            }
+            candidates.sort(Comparator.reverseOrder());
+            for (int i = 0; i < candidates.size(); i++) {
+                double a = candidates.get(i);
+                for (int j = i + 1; j < candidates.size(); j++) {
+                    double b = candidates.get(j);
+                    if (Math.abs(a + b - total) < 0.011) {
+                        if (pretax == null) pretax = round2(a);
+                        if (tax == null) tax = round2(b);
+                        break;
+                    }
+                }
+                if (tax != null && pretax != null) break;
+            }
+            // 兜底: 只剩一个候选
+            if ((tax == null || pretax == null) && candidates.size() == 1) {
+                double only = round2(candidates.get(0));
+                if (pretax == null && tax == null) {
+                    pretax = only;
+                    tax = round2(total - only);
+                } else if (tax == null) {
+                    tax = only;
+                } else if (pretax == null) {
+                    pretax = only;
+                }
+            }
+        }
+
+        // 兜底 2: total - 任一 = 另一
+        if (tax == null && total != null && pretax != null) {
+            tax = round2(total - pretax);
+        }
+        if (pretax == null && total != null && tax != null) {
+            pretax = round2(total - tax);
+        }
+
+        fields.setAmount(total);
+        fields.setTaxAmount(tax);
+        fields.setAmountPretax(pretax);
+        fields.setAmountCn(AmountUtils.extractCnAmount(norm));
+        fields.setAmountMatch(AmountUtils.amountConsistent(fields.getAmountCn(), fields.getAmount()));
+
+        // 主体
+        String sellerName, sellerTax, buyerName, buyerTax;
+        if (lines != null && !lines.isEmpty()) {
+            String[] parties = extractPartiesFromLines(lines);
+            sellerName = parties[0]; sellerTax = parties[1];
+            buyerName = parties[2]; buyerTax = parties[3];
+            if (!(sellerName != null && buyerName != null)) {
+                String[] textParties = extractPartiesFromText(norm);
+                sellerName = or(sellerName, textParties[0]);
+                sellerTax = or(sellerTax, textParties[1]);
+                buyerName = or(buyerName, textParties[2]);
+                buyerTax = or(buyerTax, textParties[3]);
+            }
+        } else {
+            String[] textParties = extractPartiesFromText(norm);
+            sellerName = textParties[0]; sellerTax = textParties[1];
+            buyerName = textParties[2]; buyerTax = textParties[3];
+        }
+        fields.setSellerName(sellerName);
+        fields.setSellerTaxNo(sellerTax);
+        fields.setBuyerName(buyerName);
+        fields.setBuyerTaxNo(buyerTax);
+
+        return fields;
+    }
+
+    private static String or(String a, String b) {
+        return a != null ? a : b;
+    }
+
+    private static double round2(double v) {
+        return Math.round(v * 100.0) / 100.0;
+    }
+
+    private static String norm(String text) {
+        return text == null ? "" : text.replaceAll("\\s+", " ").trim();
+    }
+
+    private static String detectInvoiceType(String text) {
+        for (String t : INVOICE_TYPES) {
+            if (text.contains(t)) return t;
+        }
+        return null;
+    }
+
+    private static String extractInvoiceNo(String text) {
+        Matcher m = NO_PATTERN.matcher(text);
+        return m.find() ? m.group(1) : null;
+    }
+
+    private static String extractInvoiceCode(String text) {
+        Matcher m = CODE_PATTERN.matcher(text);
+        return m.find() ? m.group(1) : null;
+    }
+
+    private static String extractDate(String text) {
+        Matcher m = DATE_PATTERN.matcher(text);
+        if (!m.find()) return null;
+        int y = Integer.parseInt(m.group(1));
+        int mo = Integer.parseInt(m.group(2));
+        int d = Integer.parseInt(m.group(3));
+        return String.format("%04d-%02d-%02d", y, mo, d);
+    }
+
+    private static String cleanName(String name) {
+        if (name == null) return null;
+        String s = name;
+        // OCR 容易把 "名称北京国钜..." 整段匹出来 (名称 是字符类里的字), 剥掉前缀保留公司名
+        for (String prefix : new String[]{"名称", "购买方", "销售方", "买方", "卖方", "购方", "销方"}) {
+            if (s.startsWith(prefix)) s = s.substring(prefix.length());
+        }
+        // 截断 stop word (label 残留: OCR/PDF 都可能把 label 跟 value 拼在一起)
+        for (String stop : new String[]{"纳税人", "统一社会", "购买方", "销售方",
+                "项目名称", "规格型号", "单价", "数量", "金额", "税率", "税额",
+                "价税合计", "大写", "小写", "备注", "收款", "复核", "开票"}) {
+            int idx = s.indexOf(stop);
+            if (idx >= 0) s = s.substring(0, idx);
+        }
+        // 去掉前导符号
+        s = s.replaceAll("^[\\s:：,，。、]+", "");
+        // 只保留中文/字母/数字/()/-/·
+        s = s.replaceAll("[^一-龥A-Za-z0-9（）()·\\-]", "");
+        s = s.replaceAll("[：:;,，。、 ]+$", "").trim();
+        return s.isEmpty() ? null : s;
+    }
+
+    /**
+     * box: [[x1,y1], ...] → (cx, cy)
+     */
+    private static double[] boxCenter(List<List<Float>> box) {
+        if (box == null || box.size() < 4) return new double[]{0, 0};
+        double minX = Double.MAX_VALUE, maxX = -Double.MAX_VALUE;
+        double minY = Double.MAX_VALUE, maxY = -Double.MAX_VALUE;
+        for (List<Float> p : box) {
+            if (p.size() < 2) continue;
+            double x = p.get(0), y = p.get(1);
+            if (x < minX) minX = x;
+            if (x > maxX) maxX = x;
+            if (y < minY) minY = y;
+            if (y > maxY) maxY = y;
+        }
+        return new double[]{(minX + maxX) / 2, (minY + maxY) / 2};
+    }
+
+    /**
+     * 返回 [seller_name, seller_tax, buyer_name, buyer_tax]
+     * <p>
+     * 策略:
+     *   - byX 分支 (左右栏): 用 buyer/seller 标签的 x 坐标分栏. **用 COMPANY_PATTERN** (不依赖"名称："前缀).
+     *   - 上下栏 / 单栏分支: 按 y 排序, **第一个 = 销售方** (中国数电票/电子专票版式 — 销方先印).
+     */
+    private String[] extractPartiesFromLines(List<OcrLine> lines) {
+        List<List<Float>> buyerLabelBox = null, sellerLabelBox = null;
+        for (OcrLine line : lines) {
+            if (buyerLabelBox == null && BUYER_LABEL.matcher(line.getText() != null ? line.getText() : "").find()) {
+                buyerLabelBox = line.getBox();
+            }
+            if (sellerLabelBox == null && SELLER_LABEL.matcher(line.getText() != null ? line.getText() : "").find()) {
+                sellerLabelBox = line.getBox();
+            }
+        }
+        boolean byX = buyerLabelBox != null && sellerLabelBox != null
+                && Math.abs(boxCenter(buyerLabelBox)[0] - boxCenter(sellerLabelBox)[0]) > 50;
+
+        String sellerName = null, sellerTax = null, buyerName = null, buyerTax = null;
+
+        if (byX) {
+            double[] bc = boxCenter(buyerLabelBox);
+            double[] sc = boxCenter(sellerLabelBox);
+            double mid = (bc[0] + sc[0]) / 2;
+            for (OcrLine line : lines) {
+                if (line.getText() == null) continue;
+                // 用 COMPANY_PATTERN 替代 NAME_PATTERN — 不依赖 "名称：" 前缀, 对拆字/拼接行更鲁棒
+                Matcher nm = COMPANY_PATTERN.matcher(line.getText());
+                if (nm.find()) {
+                    double cx = boxCenter(line.getBox())[0];
+                    String cleaned = cleanName(nm.group(1));
+                    if (cleaned == null || containsStopWord(cleaned)) continue;
+                    if (cx < mid && buyerName == null) buyerName = cleaned;
+                    else if (cx >= mid && sellerName == null) sellerName = cleaned;
+                }
+                Matcher tm = TAX_NO_PATTERN.matcher(line.getText());
+                if (tm.find()) {
+                    double cx = boxCenter(line.getBox())[0];
+                    String tax = tm.group(1);
+                    if (cx < mid && buyerTax == null) buyerTax = tax;
+                    else if (cx >= mid && sellerTax == null) sellerTax = tax;
+                }
+            }
+        } else {
+            // 没找到 buyer/seller 标签 box → 自动检测栏位 (左右栏 vs 上下栏)
+            // 收集所有 (name, cx, cy) 和 (tax, cx, cy) 候选
+            List<double[]> nameCoords = new ArrayList<>();  // [cx, cy]
+            List<String> nameVals = new ArrayList<>();
+            List<double[]> taxCoords = new ArrayList<>();
+            List<String> taxVals = new ArrayList<>();
+            for (OcrLine line : lines) {
+                if (line.getText() == null) continue;
+                double[] c = boxCenter(line.getBox());
+                Matcher nm = COMPANY_PATTERN.matcher(line.getText());
+                if (nm.find()) {
+                    String cleaned = cleanName(nm.group(1));
+                    log.info("DEBUG company match: text={} match={} cleaned={} stopWord={}",
+                            line.getText(), nm.group(1), cleaned, containsStopWord(cleaned));
+                    if (cleaned != null && !containsStopWord(cleaned)) {
+                        nameCoords.add(c);
+                        nameVals.add(cleaned);
+                    }
+                }
+                Matcher tm = TAX_NO_PATTERN.matcher(line.getText());
+                if (tm.find()) {
+                    taxCoords.add(c);
+                    taxVals.add(tm.group(1));
+                }
+            }
+
+            // 自动判断: 前两个 name 的 |dy| < 30 且 |dx| > 200 → 左右栏
+            boolean leftRight = false;
+            if (nameCoords.size() >= 2) {
+                double dx = Math.abs(nameCoords.get(0)[0] - nameCoords.get(1)[0]);
+                double dy = Math.abs(nameCoords.get(0)[1] - nameCoords.get(1)[1]);
+                leftRight = dy < 30 && dx > 200;
+            }
+
+            if (leftRight) {
+                // 左右栏: 按 x 排序, **x 小 = 买方 (购买方在左), x 大 = 卖方 (销售方在右)**
+                sortByFirstCoord(nameCoords, nameVals, 0);
+                sortByFirstCoord(taxCoords, taxVals, 0);
+            } else {
+                // 上下栏: 按 y 排序, **第一个 = 买方** (国家税务总局标准: 购方信息在前)
+                sortByFirstCoord(nameCoords, nameVals, 1);
+                sortByFirstCoord(taxCoords, taxVals, 1);
+            }
+
+            if (nameVals.size() > 0) buyerName = nameVals.get(0);
+            if (taxVals.size() > 0) buyerTax = taxVals.get(0);
+            if (nameVals.size() > 1) sellerName = nameVals.get(1);
+            if (taxVals.size() > 1) sellerTax = taxVals.get(1);
+        }
+
+        return new String[]{sellerName, sellerTax, buyerName, buyerTax};
+    }
+
+    /**
+     * 按指定坐标 (0=x, 1=y) 同步排序坐标和值列表
+     */
+    private static void sortByFirstCoord(List<double[]> coords, List<String> vals, int dim) {
+        List<Integer> idx = new ArrayList<>();
+        for (int i = 0; i < coords.size(); i++) idx.add(i);
+        idx.sort(Comparator.comparingDouble(i -> coords.get(i)[dim]));
+        List<double[]> sc = new ArrayList<>();
+        List<String> sv = new ArrayList<>();
+        for (int i : idx) {
+            sc.add(coords.get(i));
+            sv.add(vals.get(i));
+        }
+        coords.clear(); coords.addAll(sc);
+        vals.clear(); vals.addAll(sv);
+    }
+
+    /**
+     * 无 box 信息时的 fallback: 全局收集 name + tax, 按出现顺序配对.
+     * <p>
+     * 中国电子发票版式 (数电票 / 电子专票): 国家税务总局标准 — **购买方信息在前(左/上), 销售方信息在后(右/下)**.
+     * PDF 文字流的 value 顺序通常按视觉顺序 (左/上 先), 所以:
+     *   **第一个匹配 = 购买方, 第二个匹配 = 销售方**.
+     * <p>
+     * 修复:
+     *   1. PDF 内嵌文本拆字版式 ("购/买/方/信/息") — 旧 "在标签后 200 字符找 name" 失效 (label/value 分段)
+     *   2. 用 COMPANY_PATTERN (不依赖 "名称：" 前缀) 替代 NAME_PATTERN
+     *   3. 第一个 = 买方 (国家税务总局标准: 购方信息在前)
+     *
+     * @return [seller_name, seller_tax, buyer_name, buyer_tax]
+     */
+    private String[] extractPartiesFromText(String text) {
+        // 1. 全局收集所有公司名 (按出现顺序, 去重叠, 去 stop word)
+        List<String> cleanedNames = new ArrayList<>();
+        Matcher nm = COMPANY_PATTERN.matcher(text);
+        int lastEnd = -1;
+        while (nm.find()) {
+            if (nm.start() < lastEnd) continue;  // 去重叠
+            String c = cleanName(nm.group(1));
+            if (c != null && !containsStopWord(c)) cleanedNames.add(c);
+            lastEnd = nm.end();
+        }
+
+        // 2. 全局收集所有税号 (按出现顺序)
+        List<String> taxes = new ArrayList<>();
+        Matcher tm = TAX_NO_PATTERN.matcher(text);
+        while (tm.find()) taxes.add(tm.group(1));
+
+        // 3. 配对 (按位置一一对应): name[0]<->tax[0] = 买方, name[1]<->tax[1] = 卖方
+        String buyerName  = cleanedNames.size() > 0 ? cleanedNames.get(0) : null;
+        String buyerTax   = taxes.size() > 0 ? taxes.get(0) : null;
+        String sellerName = cleanedNames.size() > 1 ? cleanedNames.get(1) : null;
+        String sellerTax  = taxes.size() > 1 ? taxes.get(1) : null;
+
+        return new String[]{sellerName, sellerTax, buyerName, buyerTax};
+    }
+
+    /**
+     * 公司名 stop word 过滤 (cleanName 已经过滤一些, 这里再覆盖):
+     *   表格项目名 ("项目名称"), 规格型号 ("规格"), 备注 ("备注"), 单位, 数量 等.
+     * <p>
+     * 注意: "名称" 不在这里过滤 — cleanName 会截断 "名称" 前缀; 如果 cleanName 截断后还是包含 "名称"
+     * 才在这里过滤 (防止截断失败导致整个公司名被拒).
+     */
+    private static boolean containsStopWord(String s) {
+        if (s == null) return true;
+        // 只过滤明显不是公司名的杂质 (cleanName 已经处理过大部分 stop word)
+        return s.contains("价税合计") || s.contains("大写") || s.contains("小写")
+                || s.contains("项目名称") || s.contains("规格型号")
+                || s.equals("公司") || s.length() < 4;
+    }
+}
diff --git a/ry-api/ruoyi-business/src/main/java/com/ruoyi/business/ocr/service/QrDecoder.java b/ry-api/ruoyi-business/src/main/java/com/ruoyi/business/ocr/service/QrDecoder.java
new file mode 100644
index 0000000..2235ae4
--- /dev/null
+++ b/ry-api/ruoyi-business/src/main/java/com/ruoyi/business/ocr/service/QrDecoder.java
@@ -0,0 +1,159 @@
+package com.ruoyi.business.ocr.service;
+
+import com.google.zxing.*;
+import com.google.zxing.client.j2se.BufferedImageLuminanceSource;
+import com.google.zxing.common.HybridBinarizer;
+import com.ruoyi.business.ocr.QrDecodeResult;
+import lombok.extern.slf4j.Slf4j;
+import org.springframework.stereotype.Service;
+
+import javax.imageio.ImageIO;
+import java.awt.Graphics2D;
+import java.awt.RenderingHints;
+import java.awt.image.BufferedImage;
+import java.io.IOException;
+import java.nio.file.Path;
+
+/**
+ * 电子发票二维码识别 — 对齐 Python app.services.qr_decoder
+ * <p>
+ * 国家税务总局规范的电子发票二维码内容格式 (8 字段逗号分隔):
+ *   01,&lt;type&gt;,&lt;invoice_code&gt;,&lt;invoice_no&gt;,&lt;amount&gt;,&lt;date&gt;,&lt;check_code&gt;,&lt;reserved&gt;
+ * <p>
+ * 例: 01,31,,24922000000006110014,39500.00,20240202,,A371
+ * <p>
+ * 使用 ZXing (纯 Java, 无需 opencv).
+ */
+@Slf4j
+@Service
+public class QrDecoder {
+
+    private static final MultiFormatReader ZXING_READER = new MultiFormatReader();
+
+    /**
+     * 从图片文件解电子发票二维码
+     */
+    public QrDecodeResult decodeQr(Path imagePath) {
+        String raw;
+        try {
+            raw = detectQr(imagePath);
+        } catch (Exception e) {
+            log.warn("QR 检测异常 {}: {}", imagePath.getFileName(), e.getMessage());
+            return new QrDecodeResult();
+        }
+        if (raw == null || raw.isEmpty()) {
+            return new QrDecodeResult();
+        }
+        QrDecodeResult parsed = parsePayload(raw);
+        if (parsed.hasAnyField()) {
+            log.info("QR 解码成功: no={}, amt={}, date={}",
+                    parsed.getInvoiceNo(), parsed.getAmount(), parsed.getInvoiceDate());
+        } else {
+            log.debug("QR 解出但字段无效: raw={}", raw.substring(0, Math.min(80, raw.length())));
+        }
+        return parsed;
+    }
+
+    /**
+     * 全图 → 4 象限 → 2x 放大, 任一命中即返回.
+     */
+    private String detectQr(Path imagePath) throws IOException {
+        BufferedImage img = ImageIO.read(imagePath.toFile());
+        if (img == null) {
+            return "";
+        }
+        // 1) 全图
+        String txt = tryDecode(img);
+        if (txt != null && !txt.isEmpty()) {
+            return txt;
+        }
+        // 2) 四象限
+        int h = img.getHeight();
+        int w = img.getWidth();
+        BufferedImage[][] crops = {
+                {img.getSubimage(0, 0, w / 2, h / 2), img.getSubimage(w / 2, 0, w - w / 2, h / 2)},
+                {img.getSubimage(0, h / 2, w / 2, h - h / 2), img.getSubimage(w / 2, h / 2, w - w / 2, h - h / 2)}
+        };
+        String[] names = {"left-top", "right-top", "left-bottom", "right-bottom"};
+        int idx = 0;
+        for (BufferedImage[] row : crops) {
+            for (BufferedImage crop : row) {
+                txt = tryDecode(crop);
+                if (txt != null && !txt.isEmpty()) {
+                    log.debug("QR found in {}", names[idx]);
+                    return txt;
+                }
+                idx++;
+            }
+        }
+        // 3) 2x 放大 (二维码像素过小的情况)
+        BufferedImage scaled = resize(img, 2.0);
+        return tryDecode(scaled);
+    }
+
+    private String tryDecode(BufferedImage img) {
+        try {
+            LuminanceSource source = new BufferedImageLuminanceSource(img);
+            BinaryBitmap bitmap = new BinaryBitmap(new HybridBinarizer(source));
+            Result result = ZXING_READER.decode(bitmap);
+            return result != null ? result.getText() : "";
+        } catch (NotFoundException e) {
+            return "";
+        } catch (Exception e) {
+            log.debug("ZXing decode error: {}", e.getMessage());
+            return "";
+        }
+    }
+
+    private static BufferedImage resize(BufferedImage src, double scale) {
+        int w = (int) (src.getWidth() * scale);
+        int h = (int) (src.getHeight() * scale);
+        BufferedImage dst = new BufferedImage(w, h, BufferedImage.TYPE_INT_RGB);
+        Graphics2D g = dst.createGraphics();
+        g.setRenderingHint(RenderingHints.KEY_INTERPOLATION, RenderingHints.VALUE_INTERPOLATION_BICUBIC);
+        g.drawImage(src, 0, 0, w, h, null);
+        g.dispose();
+        return dst;
+    }
+
+    /**
+     * 解析电子发票二维码内容 → QrDecodeResult
+     * <p>
+     * 字段全 null 表示无二维码或格式不正确.
+     */
+    private QrDecodeResult parsePayload(String raw) {
+        if (raw == null || raw.isEmpty()) {
+            return new QrDecodeResult();
+        }
+        String[] parts = raw.split(",", -1);
+        if (parts.length != 8) {
+            log.debug("QR 字段数 {} != 8, 视为格式不正确", parts.length);
+            QrDecodeResult r = new QrDecodeResult();
+            r.setRaw(raw);
+            return r;
+        }
+        QrDecodeResult result = new QrDecodeResult();
+        result.setRaw(raw);
+
+        // parts[3] = 发票号
+        String invoiceNo = parts[3].trim();
+        if (!invoiceNo.isEmpty() && invoiceNo.length() >= 10 && invoiceNo.length() <= 30) {
+            result.setInvoiceNo(invoiceNo);
+        }
+
+        // parts[4] = 金额
+        String amtStr = parts[4].trim();
+        if (!amtStr.isEmpty()) {
+            try {
+                result.setAmount(Math.round(Double.parseDouble(amtStr) * 100.0) / 100.0);
+            } catch (NumberFormatException ignored) {}
+        }
+
+        // parts[5] = 开票日期 (YYYYMMDD)
+        String dateStr = parts[5].trim();
+        if (dateStr.length() == 8 && dateStr.chars().allMatch(Character::isDigit)) {
+            result.setInvoiceDate(dateStr.substring(0, 4) + "-" + dateStr.substring(4, 6) + "-" + dateStr.substring(6, 8));
+        }
+        return result;
+    }
+}
diff --git a/ry-api/ruoyi-business/src/main/java/com/ruoyi/business/ocr/service/RecognizeService.java b/ry-api/ruoyi-business/src/main/java/com/ruoyi/business/ocr/service/RecognizeService.java
new file mode 100644
index 0000000..5badece
--- /dev/null
+++ b/ry-api/ruoyi-business/src/main/java/com/ruoyi/business/ocr/service/RecognizeService.java
@@ -0,0 +1,312 @@
+package com.ruoyi.business.ocr.service;
+
+import com.ruoyi.business.ocr.config.OcrProperties;
+import com.ruoyi.business.ocr.core.ImageProcessor;
+import com.ruoyi.business.ocr.core.OcrEngine;
+import com.ruoyi.business.ocr.core.PdfProcessor;
+import com.ruoyi.business.ocr.exception.OcrTimeoutException;
+import com.ruoyi.business.ocr.InvoiceFields;
+import com.ruoyi.business.ocr.InvoiceResult;
+import com.ruoyi.business.ocr.OcrLine;
+import com.ruoyi.business.ocr.QrDecodeResult;
+import lombok.RequiredArgsConstructor;
+import lombok.extern.slf4j.Slf4j;
+import org.springframework.stereotype.Service;
+
+import java.io.IOException;
+import java.nio.file.Files;
+import java.nio.file.Path;
+import java.util.ArrayList;
+import java.util.List;
+import java.util.Set;
+
+/**
+ * 端到端识别流水: 文件 → QR → (可选)OCR → 字段抽取 → InvoiceResult
+ * <p>
+ * 完全对齐 Python app.services.recognize_service:
+ * <pre>
+ * 1. PDF / 图片 → BufferedImage 列表 (page_count)
+ * 2. qr = decode_qr(page[0])
+ * 3. qr_ok && !qr_full_ocr → [QR only] 快路径返回
+ * 4. !qr_ok → not_invoice, 不跑 OCR
+ * 5. for each page:
+ *      if elapsed > total_deadline → timeout
+ *      try _ocr_image(page) with page_timeout
+ * 6. extract_invoice(raw_text, lines)
+ * 7. overlay_qr_fields(qr)  ← QR 3 字段覆盖 OCR 结果
+ * 8. return InvoiceResult
+ * </pre>
+ */
+@Slf4j
+@Service
+@RequiredArgsConstructor
+public class RecognizeService {
+
+    private static final Set<String> IMG_EXTS = Set.of(".png", ".jpg", ".jpeg", ".bmp", ".webp", ".tif", ".tiff");
+
+    private final OcrProperties props;
+    private final PdfProcessor pdfProcessor;
+    private final OcrEngine ocrEngine;
+    private final QrDecoder qrDecoder;
+    private final InvoiceExtractor extractor;
+
+    /**
+     * 入口: 上传文件字节流
+     */
+    public InvoiceResult recognizeFile(String filename, byte[] content) {
+        long t0 = System.currentTimeMillis();
+        String suffix = filename == null ? ".bin" : ext(filename);
+        Path tmp = saveUpload(content, suffix);
+        try {
+            return recognizePath(tmp, true);
+        } finally {
+            // 已在 recognizePath finally 里删, 这里兜底
+            try { Files.deleteIfExists(tmp); } catch (IOException ignored) {}
+            log.info("总耗时 {}ms", System.currentTimeMillis() - t0);
+        }
+    }
+
+    /**
+     * 入口: 服务器本地路径 (已在白名单校验)
+     */
+    public InvoiceResult recognizePath(Path filePath, boolean deleteAfter) {
+        long t0 = System.currentTimeMillis();
+        String suffix = ext(filePath.getFileName().toString());
+        long totalDeadline = t0 + props.getOcr().getTotalTimeoutS() * 1000L;
+
+        try {
+            // 1. PDF / 图片 → 临时图片列表
+            List<Path> pageImgs;
+            try {
+                if (suffix.equals(".pdf")) {
+                    pageImgs = pdfProcessor.pdfToImages(filePath);
+                } else if (IMG_EXTS.contains(suffix)) {
+                    pageImgs = List.of(filePath);
+                } else {
+                    return err("不支持的文件类型: " + suffix + "（仅支持 PDF / 图片）",
+                            "unsupported", false, 1, t0);
+                }
+            } catch (Exception e) {
+                log.error("PDF/图片处理失败: {}", e.getMessage(), e);
+                return err("PDF/图片处理失败: " + e.getMessage(),
+                        "process_failed", false, 1, t0);
+            }
+
+            // 2. QR 优先识别
+            QrDecodeResult qr = qrDecoder.decodeQr(pageImgs.get(0));
+            boolean qrOk = qr.hasAnyField();
+
+            // 3. 快路径: QR 命中且 fast 模式
+            if (qrOk && !props.getOcr().isQrFullOcr()) {
+                log.info("QR 快路径: {} 耗时 {}ms", filePath.getFileName(), elapsedMs(t0));
+                return buildQrOnlyResult(qr, elapsedMs(t0), pageImgs.size());
+            }
+
+            // 4. 无有效 QR → 非发票, 不跑 OCR
+            if (!qrOk) {
+                String reason = qr.getRaw() == null || qr.getRaw().isEmpty()
+                        ? "未识别到发票二维码"
+                        : "二维码格式不合法";
+                log.info("非发票 (无有效 QR): {} 耗时 {}ms", filePath.getFileName(), elapsedMs(t0));
+                InvoiceResult r = err(reason + "（可能不是发票图片）",
+                        "not_invoice", false, pageImgs.size(), t0);
+                r.setFromQr(false);
+                r.setQrRaw(qr.getRaw() == null || qr.getRaw().isEmpty() ? null : qr.getRaw());
+                r.setQrError(qr.getRaw() == null || qr.getRaw().isEmpty() ? "no_qr" : "bad_format");
+                return r;
+            }
+
+            // 4.5 PDF 内嵌文本 fast path (跳过 ONNX, 直接 PDFTextStripper)
+            //    - 适用电子发票 / 数电票 PDF (含真实文本层)
+            //    - 扫描件 PDF 抽不到文本, 自动回退到 ONNX
+            if (suffix.equals(".pdf") && props.getOcr().isUsePdfTextFirst()) {
+                try {
+                    String pdfText = pdfProcessor.extractText(filePath);
+                    int minChars = props.getOcr().getPdfTextMinChars();
+                    if (pdfText.length() >= minChars) {
+                        int elapsed = elapsedMs(t0);
+                        log.info("PDF 内嵌文本 fast path: {} chars={}, 跳过 ONNX, 耗时 {}ms",
+                                filePath.getFileName(), pdfText.length(), elapsed);
+                        InvoiceFields fields = extractor.extract(pdfText, java.util.Collections.emptyList(), qr.getAmount());
+                        fields = overlayQrFields(fields, qr);
+                        InvoiceResult r = new InvoiceResult();
+                        r.setSuccess(true);
+                        r.setIsInvoice(true);
+                        r.setRawText(pdfText);
+                        r.setFields(fields);
+                        r.setPageCount(pageImgs.size());
+                        r.setEngine("pdftxt");
+                        r.setElapsedMs(elapsed);
+                        r.setFromQr(true);
+                        r.setQrRaw(qr.getRaw());
+                        return r;
+                    }
+                    log.info("PDF 内嵌文本太短 ({} 字符 < {}), 回退 ONNX", pdfText.length(), minChars);
+                } catch (Exception e) {
+                    log.warn("PDF 内嵌文本抽取失败, 回退 ONNX: {}", e.getMessage());
+                }
+            }
+
+            // 5. 每页 OCR
+            List<OcrLine> allLines = new ArrayList<>();
+            for (int idx = 0; idx < pageImgs.size(); idx++) {
+                long remaining = totalDeadline - System.currentTimeMillis();
+                if (remaining <= 0) {
+                    log.warn("达到总超时 ({}s), 中断 OCR", props.getOcr().getTotalTimeoutS());
+                    InvoiceResult r = err(String.format("达到总超时 (%d秒), 已识别 %d/%d 页",
+                                    props.getOcr().getTotalTimeoutS(), idx, pageImgs.size()),
+                            "timeout", false, pageImgs.size(), t0);
+                    r.setRawText(joinLines(allLines));
+                    r.setLines(allLines);
+                    r.setFromQr(true);
+                    r.setQrRaw(qr.getRaw());
+                    return r;
+                }
+                int pageTimeout = (int) Math.min(props.getOcr().getPageTimeoutS(), remaining / 1000.0);
+                try {
+                    allLines.addAll(ocrImage(pageImgs.get(idx)));
+                } catch (OcrTimeoutException e) {
+                    log.warn("第 {} 页 OCR 超时: {}", idx + 1, e.getMessage());
+                    InvoiceResult r = err(String.format("第 %d 页识别超时 (%.1f秒)",
+                                    idx + 1, pageTimeout * 1.0),
+                            "timeout", false, pageImgs.size(), t0);
+                    r.setRawText(joinLines(allLines));
+                    r.setLines(allLines);
+                    r.setFromQr(true);
+                    r.setQrRaw(qr.getRaw());
+                    return r;
+                } catch (Exception e) {
+                    log.error("第 {} 页 OCR 失败: {}", idx + 1, e.getMessage(), e);
+                    InvoiceResult r = err(String.format("第 %d 页识别失败: %s", idx + 1, e.getMessage()),
+                            "ocr_failed", false, pageImgs.size(), t0);
+                    r.setRawText(joinLines(allLines));
+                    r.setLines(allLines);
+                    r.setFromQr(true);
+                    r.setQrRaw(qr.getRaw());
+                    return r;
+                }
+            }
+
+            // 6. 字段抽取 (QR total 优先) + QR 字段覆盖
+            String rawText = joinLines(allLines);
+            // 把 QR 的 amount 提前传给 extractor, 让 fallback 组合搜索能用对的总价.
+            InvoiceFields fields = extractor.extract(rawText, allLines, qr.getAmount());
+            fields = overlayQrFields(fields, qr);
+
+            int elapsed = elapsedMs(t0);
+            log.info("识别完成: {} 页={}, from_qr=true, 耗时={}ms", filePath.getFileName(), pageImgs.size(), elapsed);
+            InvoiceResult r = new InvoiceResult();
+            r.setSuccess(true);
+            r.setIsInvoice(true);
+            r.setRawText(rawText);
+            r.setLines(allLines);
+            r.setFields(fields);
+            r.setPageCount(pageImgs.size());
+            r.setEngine("paddleocr");
+            r.setElapsedMs(elapsed);
+            r.setFromQr(true);
+            r.setQrRaw(qr.getRaw());
+            return r;
+
+        } finally {
+            if (deleteAfter) {
+                try {
+                    Files.deleteIfExists(filePath);
+                    if (suffix.equals(".pdf")) {
+                        // 清理 .xxx_pages/ 临时目录
+                        Path pagesDir = filePath.getParent().resolve("." + stem(filePath) + "_pages");
+                        if (Files.exists(pagesDir)) {
+                            try (var stream = Files.walk(pagesDir)) {
+                                stream.sorted((a, b) -> b.compareTo(a)).forEach(p -> {
+                                    try { Files.deleteIfExists(p); } catch (IOException ignored) {}
+                                });
+                            }
+                        }
+                    }
+                } catch (Exception ignored) {}
+            }
+        }
+    }
+
+    // ---------- 内部 ----------
+
+    private List<OcrLine> ocrImage(Path imgPath) {
+        Path rotated = ImageProcessor.autoRotate(imgPath);
+        Path enhanced = ImageProcessor.enhance(rotated);
+        return ocrEngine.recognize(enhanced);
+    }
+
+    private static String joinLines(List<OcrLine> lines) {
+        StringBuilder sb = new StringBuilder();
+        for (int i = 0; i < lines.size(); i++) {
+            if (i > 0) sb.append('\n');
+            sb.append(lines.get(i).getText());
+        }
+        return sb.toString();
+    }
+
+    private static String ext(String filename) {
+        int dot = filename.lastIndexOf('.');
+        return dot >= 0 ? filename.substring(dot).toLowerCase() : "";
+    }
+
+    private static String stem(Path p) {
+        String name = p.getFileName().toString();
+        int dot = name.lastIndexOf('.');
+        return dot > 0 ? name.substring(0, dot) : name;
+    }
+
+    private static Path saveUpload(byte[] content, String suffix) {
+        try {
+            Path tmp = Files.createTempFile("ry_ocr_", suffix);
+            Files.write(tmp, content);
+            return tmp;
+        } catch (IOException e) {
+            throw new RuntimeException("保存临时文件失败: " + e.getMessage(), e);
+        }
+    }
+
+    private static int elapsedMs(long t0) {
+        return (int) (System.currentTimeMillis() - t0);
+    }
+
+    private InvoiceResult err(String error, String errorCode, boolean isInvoice, int pageCount, long t0) {
+        InvoiceResult r = new InvoiceResult();
+        r.setSuccess(false);
+        r.setIsInvoice(isInvoice);
+        r.setError(error);
+        r.setErrorCode(errorCode);
+        r.setPageCount(pageCount);
+        r.setEngine("paddleocr");
+        r.setElapsedMs(elapsedMs(t0));
+        return r;
+    }
+
+    private InvoiceResult buildQrOnlyResult(QrDecodeResult qr, int elapsedMs, int pageCount) {
+        InvoiceFields fields = new InvoiceFields();
+        fields.setInvoiceNo(qr.getInvoiceNo());
+        fields.setAmount(qr.getAmount());
+        fields.setInvoiceDate(qr.getInvoiceDate());
+        InvoiceResult r = new InvoiceResult();
+        r.setSuccess(true);
+        r.setIsInvoice(true);
+        r.setRawText("[QR only] " + qr.getRaw());
+        r.setFields(fields);
+        r.setPageCount(pageCount);
+        r.setEngine("qr");
+        r.setElapsedMs(elapsedMs);
+        r.setFromQr(true);
+        r.setQrRaw(qr.getRaw());
+        return r;
+    }
+
+    /**
+     * QR 解出的 3 字段优先, 没解到的保持 OCR 结果
+     */
+    private InvoiceFields overlayQrFields(InvoiceFields fields, QrDecodeResult qr) {
+        if (qr.getInvoiceNo() != null) fields.setInvoiceNo(qr.getInvoiceNo());
+        if (qr.getAmount() != null) fields.setAmount(qr.getAmount());
+        if (qr.getInvoiceDate() != null) fields.setInvoiceDate(qr.getInvoiceDate());
+        return fields;
+    }
+}
diff --git a/ry-api/ruoyi-business/src/main/java/com/ruoyi/business/ocr/util/AmountUtils.java b/ry-api/ruoyi-business/src/main/java/com/ruoyi/business/ocr/util/AmountUtils.java
new file mode 100644
index 0000000..71334c6
--- /dev/null
+++ b/ry-api/ruoyi-business/src/main/java/com/ruoyi/business/ocr/util/AmountUtils.java
@@ -0,0 +1,296 @@
+package com.ruoyi.business.ocr.util;
+
+import java.util.HashMap;
+import java.util.Map;
+import java.util.regex.Matcher;
+import java.util.regex.Pattern;
+
+/**
+ * 金额工具: 中文大写金额解析 + 小写金额正则 — 对齐 Python app.utils.amount_utils
+ * <p>
+ * 中文大写金额解析是 cn2an smart 模式的简化版:
+ *   - 支持字符: 零壹贰叁肆伍陆柒捌玖拾佰仟万亿圆元角分整
+ *   - 处理 亿/万/元 三段累计, 段内 仟佰拾 累加
+ *   - 处理 角分 (0.1 / 0.01)
+ */
+public final class AmountUtils {
+
+    private AmountUtils() {}
+
+    // ---------- 小写金额正则 ----------
+
+    /** 通用金额: ¥1,234.56 或 1234.56 */
+    private static final Pattern NUM_PATTERN = Pattern.compile("¥?\\s*(\\d{1,8}(?:,\\d{3})*\\.\\d{2})");
+
+    /** 税额: "税额 ¥12.34" / "税额:12.34" */
+    private static final Pattern TAX_AMOUNT_PATTERN = Pattern.compile("税\\s*额\\s*[¥:]?\\s*(\\d{1,8}(?:,\\d{3})*\\.\\d{2})");
+
+    /** 不含税: "不含税价 ¥1234.56" / "不含税:1234.56" */
+    private static final Pattern PRETAX_PATTERN = Pattern.compile("(?:不合?税价|不含税)\\s*[¥:]?\\s*(\\d{1,8}(?:,\\d{3})*\\.\\d{2})");
+
+    /** 价税合计: "价税合计 ¥1234.56" */
+    private static final Pattern TOTAL_PATTERN = Pattern.compile("价税合计[^\\d]*[¥]?\\s*(\\d{1,8}(?:,\\d{3})*\\.\\d{2})");
+
+    // ---------- 中文大写金额正则 ----------
+
+    private static final Pattern CN_IN_PARENS_AFTER_TOTAL = Pattern.compile(
+            "价税合计[^\\(（]*[\\(（]([零壹贰叁肆伍陆柒捌玖拾佰仟万亿圆元角分整]+)[\\)）]"
+    );
+
+    private static final Pattern CN_IN_ANY_PARENS = Pattern.compile(
+            "[\\(（]([零壹贰叁肆伍陆柒捌玖拾佰仟万亿圆元角分整]{3,30})[\\)）]"
+    );
+
+    private static final Pattern CN_LONG = Pattern.compile("[零壹贰叁肆伍陆柒捌玖拾佰仟万亿圆元角分整]{3,30}");
+
+    // ---------- 数字映射 ----------
+
+    private static final Map<Character, Integer> DIGIT_MAP = new HashMap<>();
+    private static final Map<Character, Double> UNIT_MAP = new HashMap<>();
+    private static final Map<Character, Double> BIG_UNIT_MAP = new HashMap<>();
+
+    static {
+        DIGIT_MAP.put('零', 0);
+        DIGIT_MAP.put('壹', 1);
+        DIGIT_MAP.put('贰', 2);
+        DIGIT_MAP.put('叁', 3);
+        DIGIT_MAP.put('肆', 4);
+        DIGIT_MAP.put('伍', 5);
+        DIGIT_MAP.put('陆', 6);
+        DIGIT_MAP.put('柒', 7);
+        DIGIT_MAP.put('捌', 8);
+        DIGIT_MAP.put('玖', 9);
+
+        UNIT_MAP.put('拾', 10.0);
+        UNIT_MAP.put('佰', 100.0);
+        UNIT_MAP.put('仟', 1000.0);
+
+        BIG_UNIT_MAP.put('角', 0.1);
+        BIG_UNIT_MAP.put('分', 0.01);
+    }
+
+    // ---------- 公开方法 ----------
+
+    /**
+     * 从文本里抽取中文大写金额
+     * <p>
+     * 优先级:
+     *   1. "价税合计" 后面括号内
+     *   2. 任意中括号里的中文金额
+     *   3. 含"元"或"圆"的最长中文字符串
+     */
+    public static String extractCnAmount(String text) {
+        if (text == null || text.isEmpty()) {
+            return null;
+        }
+
+        Matcher m = CN_IN_PARENS_AFTER_TOTAL.matcher(text);
+        if (m.find()) {
+            return m.group(1);
+        }
+
+        m = CN_IN_ANY_PARENS.matcher(text);
+        if (m.find()) {
+            return m.group(1);
+        }
+
+        Matcher m2 = CN_LONG.matcher(text);
+        while (m2.find()) {
+            String cand = m2.group();
+            if (cand.contains("元") || cand.contains("圆")) {
+                return cand;
+            }
+        }
+        return null;
+    }
+
+    /**
+     * 中文大写金额 → float, 例如 "贰佰元整" → 200.0
+     */
+    public static Double parseCnAmount(String cnText) {
+        if (cnText == null || cnText.isEmpty()) {
+            return null;
+        }
+        try {
+            String s = normalizeCnAmount(cnText);
+            s = s.replaceAll("整$", "");
+            if (!s.endsWith("元")) {
+                s = s + "元";
+            }
+            return smartParse(s);
+        } catch (Exception e) {
+            return null;
+        }
+    }
+
+    /**
+     * 大写 vs 小写金额比对
+     */
+    public static Boolean amountConsistent(String cnText, Double numAmount) {
+        if (cnText == null || numAmount == null) {
+            return null;
+        }
+        Double cnValue = parseCnAmount(cnText);
+        if (cnValue == null) {
+            return null;
+        }
+        return Math.abs(cnValue - numAmount) < 0.011;
+    }
+
+    /**
+     * 抽取第一个形如 1234.56 或 ¥1,234.56 的金额
+     */
+    public static Double extractNumAmount(String text) {
+        if (text == null || text.isEmpty()) {
+            return null;
+        }
+        Matcher m = NUM_PATTERN.matcher(text);
+        if (!m.find()) {
+            return null;
+        }
+        try {
+            return Double.parseDouble(m.group(1).replace(",", ""));
+        } catch (NumberFormatException e) {
+            return null;
+        }
+    }
+
+    /**
+     * 抽取价税合计 (优先), 兜底走 extractNumAmount
+     */
+    public static Double extractTotalAmount(String text) {
+        if (text == null || text.isEmpty()) {
+            return null;
+        }
+        Matcher m = TOTAL_PATTERN.matcher(text);
+        if (m.find()) {
+            try {
+                return Double.parseDouble(m.group(1).replace(",", ""));
+            } catch (NumberFormatException ignored) {}
+        }
+        return extractNumAmount(text);
+    }
+
+    /**
+     * 抽取税额
+     */
+    public static Double extractTaxAmount(String text) {
+        if (text == null || text.isEmpty()) {
+            return null;
+        }
+        Matcher m = TAX_AMOUNT_PATTERN.matcher(text);
+        if (m.find()) {
+            try {
+                return Double.parseDouble(m.group(1).replace(",", ""));
+            } catch (NumberFormatException ignored) {}
+        }
+        return null;
+    }
+
+    /**
+     * 抽取不含税金额
+     */
+    public static Double extractPretaxAmount(String text) {
+        if (text == null || text.isEmpty()) {
+            return null;
+        }
+        Matcher m = PRETAX_PATTERN.matcher(text);
+        if (m.find()) {
+            try {
+                return Double.parseDouble(m.group(1).replace(",", ""));
+            } catch (NumberFormatException ignored) {}
+        }
+        return null;
+    }
+
+    // ---------- 内部 ----------
+
+    /**
+     * 中文金额归一化: 圆→元, 〇→零, 去空格
+     */
+    private static String normalizeCnAmount(String text) {
+        if (text == null) return "";
+        return text.replace("圆", "元").replace("〇", "零").replace(" ", "");
+    }
+
+    /**
+     * cn2an smart 模式简化实现:
+     *   - 拾佰仟 在元段内累加
+     *   - 万 / 亿 切换大段
+     *   - 角分 处理小数
+     */
+    private static double smartParse(String s) {
+        // 拆分: 整数部分 (元段) + 小数部分 (角分)
+        int yuanIdx = s.indexOf('元');
+        String intPart = yuanIdx >= 0 ? s.substring(0, yuanIdx) : s;
+        String decPart = yuanIdx >= 0 ? s.substring(yuanIdx + 1) : "";
+
+        double intValue = parseIntegerPart(intPart);
+        double decValue = parseDecimalPart(decPart);
+        return intValue + decValue;
+    }
+
+    /**
+     * 解析整数部分: 处理 拾佰仟 万 亿
+     */
+    private static double parseIntegerPart(String s) {
+        if (s == null || s.isEmpty()) return 0;
+
+        // 分段: 以 亿 / 万 分隔
+        // 例: 叁万玖仟伍佰 → [叁] (亿段=0) + [玖] (万段) + [伍佰] (元段)
+        // 例: 壹亿贰仟万叁仟 → [壹] (亿段) + [贰] (万段) + [叁] (元段)
+        double total = 0;
+        double currentSection = 0;  // 当前段(元/万/亿) 的累加值
+        double currentNum = 0;      // 当前数字 (0-9)
+
+        for (int i = 0; i < s.length(); i++) {
+            char c = s.charAt(i);
+            if (DIGIT_MAP.containsKey(c)) {
+                currentNum = DIGIT_MAP.get(c);
+            } else if (UNIT_MAP.containsKey(c)) {
+                // 拾佰仟: 处理 拾伍 = 15 (省略壹) 的情况
+                double v = currentNum == 0 ? 1 : currentNum;
+                currentSection += v * UNIT_MAP.get(c);
+                currentNum = 0;
+            } else if (c == '万') {
+                currentSection += currentNum;
+                total += currentSection * 10000;
+                currentSection = 0;
+                currentNum = 0;
+            } else if (c == '亿') {
+                currentSection += currentNum;
+                total += currentSection * 100000000;
+                currentSection = 0;
+                currentNum = 0;
+            }
+            // 零 跳过
+        }
+        // 收尾: 段尾若有数字未乘单位 (如 "叁万玖" 末尾的 "玖")
+        currentSection += currentNum;
+        total += currentSection;
+        return total;
+    }
+
+    /**
+     * 解析小数部分: 角分
+     */
+    private static double parseDecimalPart(String s) {
+        if (s == null || s.isEmpty()) return 0;
+
+        double value = 0;
+        double currentNum = 0;
+        for (int i = 0; i < s.length(); i++) {
+            char c = s.charAt(i);
+            if (DIGIT_MAP.containsKey(c)) {
+                currentNum = DIGIT_MAP.get(c);
+            } else if (BIG_UNIT_MAP.containsKey(c)) {
+                if (currentNum > 0 || i == 0) {
+                    double v = currentNum == 0 ? 1 : currentNum;
+                    value += v * BIG_UNIT_MAP.get(c);
+                }
+                currentNum = 0;
+            }
+        }
+        return value;
+    }
+}
diff --git a/ry-api/ruoyi-business/src/main/java/com/ruoyi/business/service/impl/InvoiceOcrService.java b/ry-api/ruoyi-business/src/main/java/com/ruoyi/business/service/impl/InvoiceOcrService.java
index d79933f..0f881d4 100644
--- a/ry-api/ruoyi-business/src/main/java/com/ruoyi/business/service/impl/InvoiceOcrService.java
+++ b/ry-api/ruoyi-business/src/main/java/com/ruoyi/business/service/impl/InvoiceOcrService.java
@@ -17,7 +17,7 @@ import com.ruoyi.business.domain.BizMeetingInvoice;
 import com.ruoyi.business.mapper.BizMeetingInvoiceMapper;
 import com.ruoyi.business.ocr.InvoiceFields;
 import com.ruoyi.business.ocr.InvoiceResult;
-import com.ruoyi.business.ocr.OcrClient;
+import com.ruoyi.business.ocr.LocalInvoiceRecognizer;
 import com.ruoyi.business.ocr.ZipExtractor;
 import com.ruoyi.business.oss.OssUploader;
 import com.ruoyi.common.utils.SecurityUtils;
@@ -44,7 +44,7 @@ public class InvoiceOcrService
     private static final Logger log = LoggerFactory.getLogger(InvoiceOcrService.class);
 
     @Autowired
-    private OcrClient ocrClient;
+    private LocalInvoiceRecognizer recognizer;
 
     @Autowired
     private OssUploader ossUploader;
@@ -159,7 +159,7 @@ public class InvoiceOcrService
      */
     void recognizeSingle(Long materialId, String ossUrl)
     {
-        InvoiceResult ir = ocrClient.recognizeByUrl(ossUrl);
+        InvoiceResult ir = recognizer.recognizeByUrl(ossUrl);
         if (Boolean.TRUE.equals(ir.getSuccess()) && ir.getFields() != null && isRecognizedAsInvoice(ir.getFields()))
         {
             BigDecimal amount = ir.getFields().getAmount() != null
@@ -210,7 +210,7 @@ public class InvoiceOcrService
             {
                 try
                 {
-                    InvoiceResult ir = ocrClient.recognize(f);
+                    InvoiceResult ir = recognizer.recognize(f);
                     if (!Boolean.TRUE.equals(ir.getSuccess()) || ir.getFields() == null
                         || !isRecognizedAsInvoice(ir.getFields()))
                     {
@@ -275,7 +275,7 @@ public class InvoiceOcrService
         if (inv == null || inv.getOssUrl() == null) return;
         try
         {
-            InvoiceResult ir = ocrClient.recognizeByUrl(inv.getOssUrl());
+            InvoiceResult ir = recognizer.recognizeByUrl(inv.getOssUrl());
             if (Boolean.TRUE.equals(ir.getSuccess()) && ir.getFields() != null && isRecognizedAsInvoice(ir.getFields()))
             {
                 BigDecimal amount = ir.getFields().getAmount() != null
diff --git a/ry-api/ruoyi-business/src/main/java/com/ruoyi/business/sms/AliyunSmsSender.java b/ry-api/ruoyi-business/src/main/java/com/ruoyi/business/sms/AliyunSmsSender.java
index 30d002f..4da25d6 100644
--- a/ry-api/ruoyi-business/src/main/java/com/ruoyi/business/sms/AliyunSmsSender.java
+++ b/ry-api/ruoyi-business/src/main/java/com/ruoyi/business/sms/AliyunSmsSender.java
@@ -45,7 +45,7 @@ public class AliyunSmsSender {
     @Value("${ruoyi.sms.esignTemplate}")
     private String esignTemplate;
 
-    @Value("${ruoyi.sms.esignBaseUrl:https://ringdoctor.com/hg}")
+    @Value("${ruoyi.sms.esignBaseUrl:https://risingdoctor.com/hg}")
     private String esignBaseUrl;
 
     @Value("${ruoyi.sms.regionId:cn-hangzhou}")
@@ -105,7 +105,7 @@ public class AliyunSmsSender {
 
     /**
      * 拼电子签签署链接: {esignBaseUrl}/#/doctor/sign-fill?attendeeId={attendeeId}
-     * 例: https://ringdoctor.com/hg/#/doctor/sign-fill?attendeeId=123
+     * 例: https://risingdoctor.com/hg/#/doctor/sign-fill?attendeeId=123
      * (nginx 子路径 /hg 已配在 ruoyi.sms.esignBaseUrl 里, 前端 Vue Router 是 hash 模式,
      *  所以 Java 只拼 #/doctor/sign-fill 路由 + attendeeId)
      */
diff --git a/ry-h5/composables/useCamera.ts b/ry-h5/composables/useCamera.ts
index 0d4b0a7..a1776dd 100644
--- a/ry-h5/composables/useCamera.ts
+++ b/ry-h5/composables/useCamera.ts
@@ -32,7 +32,27 @@ export function useCamera(videoId: string, blur?: BlurConfig) {
 	const errorMsg = ref<string>('')
 
 	function getVideoEl(): HTMLVideoElement | null {
-		return document.getElementById(videoId) as HTMLVideoElement | null
+		const el = document.getElementById(videoId) as HTMLElement | null
+		if (!el) return null
+		// uni-app H5 把 <video> 编译成 <uni-video> Vue 组件时, getElementById
+		// 拿到的是 wrapper, 它的 srcObject setter 会透传给内部原生 <video>,
+		// 但 readyState / play() 不一定透传. 用 readyState 是否 number 判断是否原生.
+		if (el.tagName === 'VIDEO' && typeof (el as any).readyState === 'number') {
+			return el as HTMLVideoElement
+		}
+		// wrapper 场景: 内部包了一个原生 <video>
+		const inner = el.querySelector('video') as HTMLVideoElement | null
+		if (inner && typeof inner.readyState === 'number') {
+			console.log('[camera] 穿透 uni-app wrapper, 找到内部原生 <video>')
+			return inner
+		}
+		// 最后一搏: 找页面上第一个原生 <video>
+		const fallback = document.querySelector('video') as HTMLVideoElement | null
+		if (fallback) {
+			console.log('[camera] 全局兜底找到第一个原生 <video>')
+			return fallback
+		}
+		return null
 	}
 
 	async function startCamera() {
diff --git a/ry-h5/pages/panorama/index.uvue b/ry-h5/pages/panorama/index.uvue
index afa3f90..2022e1d 100644
--- a/ry-h5/pages/panorama/index.uvue
+++ b/ry-h5/pages/panorama/index.uvue
@@ -14,7 +14,6 @@
 			<video
 				id="panorama-video"
 				class="video"
-				autoplay
 				muted
 				playsinline
 			/>
@@ -279,19 +278,38 @@ function goBack() {
 	.preview-actions {
 		display: flex;
 		flex-direction: row;
-		justify-content: space-around;
+		justify-content: space-between;
+		align-items: center;
+		box-sizing: border-box;
 		width: 100%;
 		padding: 40rpx 60rpx;
 		padding-bottom: calc(40rpx + env(safe-area-inset-bottom));
 	}
 
 	.preview-btn {
-		flex: 1;
-		margin: 0 20rpx;
-		padding: 24rpx 0;
+		/* flex item: 在 .preview-actions 里 flex:1 平分宽度 */
+		flex: 1 1 0%;
+		min-width: 0;
+		box-sizing: border-box;
+
+		/* flex container: 让内部 <text> 真正居中 */
+		display: flex;
+		justify-content: center;
+		align-items: center;
+
+		/* 重置浏览器 / uni-app H5 给 <button> 的默认外观 (边框/appearance 是按钮不对称的常见根因) */
+		margin: 0 16rpx;
+		padding: 0;
+		border: none;
+		outline: none;
+		appearance: none;
+		-webkit-appearance: none;
+
 		border-radius: 48rpx;
 		font-size: 30rpx;
 		font-weight: bold;
+		line-height: 1;
+		min-height: 88rpx;
 	}
 
 	.preview-btn-cancel {
diff --git a/ry-h5/pages/signin/index.uvue b/ry-h5/pages/signin/index.uvue
index 8d57088..3be623a 100644
--- a/ry-h5/pages/signin/index.uvue
+++ b/ry-h5/pages/signin/index.uvue
@@ -14,7 +14,6 @@
 			<video
 				id="signin-video"
 				class="video"
-				autoplay
 				muted
 				playsinline
 			/>
@@ -67,7 +66,7 @@
 import { useCamera } from '@/composables/useCamera'
 
 const videoId = 'signin-video'
-// A4 纸 20%~50% 区间 (签到人手机号/身份证号所在的姓名栏) 做高斯模糊脱敏
+// A4 纸: 从顶边到底边的一竖条, 水平位置在 20%~50% 宽度 (签到人手机号/身份证号所在列) 做高斯模糊脱敏
 const { captured, errorMsg, startCamera, flipCamera, takePhoto, retake, confirm } =
 	useCamera(videoId, { start: 0.2, end: 0.5, radius: 20 })
 
@@ -273,19 +272,38 @@ function goBack() {
 	.preview-actions {
 		display: flex;
 		flex-direction: row;
-		justify-content: space-around;
+		justify-content: space-between;
+		align-items: center;
+		box-sizing: border-box;
 		width: 100%;
 		padding: 40rpx 60rpx;
 		padding-bottom: calc(40rpx + env(safe-area-inset-bottom));
 	}
 
 	.preview-btn {
-		flex: 1;
-		margin: 0 20rpx;
-		padding: 24rpx 0;
+		/* flex item: 在 .preview-actions 里 flex:1 平分宽度 */
+		flex: 1 1 0%;
+		min-width: 0;
+		box-sizing: border-box;
+
+		/* flex container: 让内部 <text> 真正居中 */
+		display: flex;
+		justify-content: center;
+		align-items: center;
+
+		/* 重置浏览器 / uni-app H5 给 <button> 的默认外观 (边框/appearance 是按钮不对称的常见根因) */
+		margin: 0 16rpx;
+		padding: 0;
+		border: none;
+		outline: none;
+		appearance: none;
+		-webkit-appearance: none;
+
 		border-radius: 48rpx;
 		font-size: 30rpx;
 		font-weight: bold;
+		line-height: 1;
+		min-height: 88rpx;
 	}
 
 	.preview-btn-cancel {
diff --git a/ry-h5/utils/camera.ts b/ry-h5/utils/camera.ts
index 578e4da..4919e3b 100644
--- a/ry-h5/utils/camera.ts
+++ b/ry-h5/utils/camera.ts
@@ -8,6 +8,25 @@
  */
 export type Facing = 'user' | 'environment'
 
+// uni-app H5 runtime (uni-h5.es.js) 看到 <video autoplay> 会抢先调一次 play(),
+// 此时 srcObject 还没设,会抛 NotSupportedError 变成 unhandled rejection.
+// 模板已不再写 autoplay, 这里再装一个兜底监听, 把漏网的 NotSupportedError /
+// AbortError 静默掉, 不污染控制台.
+declare global {
+	interface Window {
+		__cameraRejectionGuard?: boolean
+	}
+}
+if (typeof window !== 'undefined' && !window.__cameraRejectionGuard) {
+	window.__cameraRejectionGuard = true
+	window.addEventListener('unhandledrejection', (e) => {
+		const name = (e?.reason as DOMException)?.name
+		if (name === 'NotSupportedError' || name === 'AbortError') {
+			e.preventDefault()
+		}
+	})
+}
+
 export async function openCamera(video: HTMLVideoElement, facing: Facing): Promise<void> {
 	if (!navigator.mediaDevices?.getUserMedia) {
 		throw new Error('当前浏览器不支持摄像头访问,请升级浏览器或使用 Chrome / Safari')
@@ -52,10 +71,56 @@ video.setAttribute('playsinline', 'true')
 // iOS 必须: 不静音黑屏 (否则 iOS 拒绝播放)
 video.muted = true
 
-// 显式 play(): <video autoplay> 只在元素首次加载时触发一次, 而 srcObject 是
-// 异步挂上的, 此时视频还没源, 浏览器不会自动重放 → 黑屏 + 播放按钮.
-// 静音 + playsinline 下, 现代浏览器默认放行静音自动播放, 直接 play() 起流.
-video.play()
+// 源就绪后再 play(): 过早 play() 会因为 video 还没有源抛 NotSupportedError
+// ("The element has no supported sources"), 变成未捕获的 Promise 拒绝.
+// safePlay 现在做四件事:
+//   1. 防御 uni-app H5 包装层让 video.play 返回 undefined 的情况
+//   2. 吞掉 NotSupportedError / AbortError (浏览器认为没源时不报)
+//   3. readyState < 2 时按 readyState 升级主动重试, 不再只等 loadedmetadata
+//   4. 其他错误 warn 后不再重试 (避免噪音)
+const safePlay = () => {
+	// 防御: uni-app H5 包装层可能让 video.play 不存在或返回 undefined
+	if (typeof video.play !== 'function') {
+		console.warn('[camera] video.play 不是函数 (可能 uni-app 包装), 跳过')
+		return
+	}
+	let result: unknown
+	try {
+		result = video.play()
+	} catch (e) {
+		console.warn('[camera] play() 同步抛错:', (e as Error)?.message)
+		return
+	}
+	if (!result || typeof (result as Promise<void>).catch !== 'function') {
+		// play() 没返回 Promise (uni-app 包装层常见), 无错误可处理, 直接放过
+		return
+	}
+	;(result as Promise<void>).catch((e: DOMException) => {
+		const name = e?.name
+		if (name === 'NotSupportedError' || name === 'AbortError') {
+			// 源还没就绪 (readyState 0/1), 起一个轮询, 数据到位再 play
+			if (video.readyState < 2) {
+				console.log('[camera] 源未就绪, readyState=', video.readyState, ', 等待后重试')
+				const start = Date.now()
+				const timer = setInterval(() => {
+					if (video.readyState >= 2) {
+						clearInterval(timer)
+						safePlay()
+					} else if (Date.now() - start > 3000) {
+						clearInterval(timer)
+						console.warn('[camera] 等待 readyState>=2 超时, 放弃自动播放')
+					}
+				}, 100)
+			}
+		} else {
+			console.warn('[camera] play() 失败:', name, e?.message)
+		}
+	})
+}
+
+// 赋值 srcObject 立刻主动 play 一次 (不再只等 loadedmetadata, 避免 uni-app
+// 包装或某些浏览器不触发该事件时视频永远不播)
+safePlay()
 
 const fallbackTimer = setTimeout(() => {
 	if (video.readyState < 1) {
@@ -67,6 +132,8 @@ const fallbackTimer = setTimeout(() => {
 				configurable: true,
 			})
 			console.log('[camera] defineProperty 后 srcObject isStream:', video.srcObject === stream)
+			// 强制赋值后再主动试一次 play
+			safePlay()
 		} catch (e) {
 			console.error('[camera] defineProperty 也失败:', e)
 		}
@@ -75,6 +142,7 @@ const fallbackTimer = setTimeout(() => {
 			console.warn('[camera] 仍 readyState 0,改用 src=blobURL')
 			try {
 				video.src = URL.createObjectURL(stream)
+				safePlay()
 			} catch (e) {
 				console.error('[camera] blobURL 兜底失败:', e)
 			}
@@ -87,11 +155,19 @@ video.addEventListener(
 	() => {
 		clearTimeout(fallbackTimer)
 		console.log('[camera] loadedmetadata 触发, readyState:', video.readyState, 'isStream:', video.srcObject === stream)
-		// 数据真正就绪后再补一次 play(), 兜底个别浏览器首次 play() 因无数据被打断
-		video.play()
+		// 数据真正就绪后再 play
+		safePlay()
 	},
 	{ once: true }
 )
+
+// 兜底: 个别浏览器 loadedmetadata 后 readyState 仍不够, loadeddata 再补一次 play()
+video.addEventListener('loadeddata', () => safePlay(), { once: true })
+
+// 若赋值 srcObject 前已就绪 (理论上不会), 立即放
+if (video.readyState >= 1) {
+	safePlay()
+}
 }
 
 export function stopCamera(video: HTMLVideoElement): void {
@@ -134,30 +210,27 @@ function buildGaussianKernel(radius: number): number[] {
 }
 
 /**
- * 对 RGBA 数组的竖直条带 [y0, y1) 做可分离高斯模糊, 返回新数组 (不修改原数组).
- * 水平卷积范围向外扩 radius, 保证条带边界处垂直卷积有正确的邻域上下文.
+ * 对 RGBA 数组的竖直条带 [x0, x1) 做可分离高斯模糊, 返回新数组 (不修改原数组).
+ * 条带贯穿全高 (顶边→底边), 只在水平 [x0, x1) 范围内模糊; 其余区域照抄.
  */
 function gaussianBlurBand(
 	data: Uint8ClampedArray,
 	width: number,
 	height: number,
-	y0: number,
-	y1: number,
+	x0: number,
+	x1: number,
 	radius: number
 ): Uint8ClampedArray {
 	const kernel = buildGaussianKernel(radius)
 	const r = radius
 	const kSize = kernel.length
 
-	const hy0 = Math.max(0, y0 - r)
-	const hy1 = Math.min(height, y1 + r)
-
-	// 1) 水平卷积: 读 data, 写 h1 (仅 [hy0, hy1) 行, 其余照抄)
+	// 1) 水平卷积 (沿 x): 读 data, 写 h1 (仅 [x0, x1) 列, 全高, 其余照抄)
 	const h1 = new Uint8ClampedArray(data)
-	const row = new Array<number>(width * 4)
-	for (let y = hy0; y < hy1; y++) {
-		const base = y * width * 4
-		for (let x = 0; x < width; x++) {
+	const col = new Array<number>(height * 4)
+	for (let x = x0; x < x1; x++) {
+		for (let y = 0; y < height; y++) {
+			const base = (y * width + x) * 4
 			let rr = 0
 			let gg = 0
 			let bb = 0
@@ -166,29 +239,34 @@ function gaussianBlurBand(
 				let sx = x + k - r
 				if (sx < 0) sx = 0
 				if (sx >= width) sx = width - 1
-				const idx = base + sx * 4
+				const idx = (y * width + sx) * 4
 				const wt = kernel[k]
 				rr += data[idx] * wt
 				gg += data[idx + 1] * wt
 				bb += data[idx + 2] * wt
 				aa += data[idx + 3] * wt
 			}
-			const o = x * 4
-			row[o] = rr
-			row[o + 1] = gg
-			row[o + 2] = bb
-			row[o + 3] = aa
+			const o = y * 4
+			col[o] = rr
+			col[o + 1] = gg
+			col[o + 2] = bb
+			col[o + 3] = aa
+		}
+		for (let y = 0; y < height; y++) {
+			const idx = (y * width + x) * 4
+			const o = y * 4
+			h1[idx] = col[o]
+			h1[idx + 1] = col[o + 1]
+			h1[idx + 2] = col[o + 2]
+			h1[idx + 3] = col[o + 3]
 		}
-		for (let i = 0; i < width * 4; i++) h1[base + i] = row[i]
 	}
 
-	// 2) 垂直卷积: 读 h1, 写 h2 (仅 [y0, y1) 行, 其余照抄 h1)
+	// 2) 垂直卷积 (沿 y): 读 h1, 写 h2 (仅 [x0, x1) 列, 全高, 其余照抄 h1)
 	const h2 = new Uint8ClampedArray(h1)
-	const band = y1 - y0
-	const col = new Array<number>(band * 4)
-	for (let x = 0; x < width; x++) {
-		for (let yy = 0; yy < band; yy++) {
-			const y = y0 + yy
+	const col2 = new Array<number>(height * 4)
+	for (let x = x0; x < x1; x++) {
+		for (let y = 0; y < height; y++) {
 			let rr = 0
 			let gg = 0
 			let bb = 0
@@ -204,19 +282,19 @@ function gaussianBlurBand(
 				bb += h1[idx + 2] * wt
 				aa += h1[idx + 3] * wt
 			}
-			const o = yy * 4
-			col[o] = rr
-			col[o + 1] = gg
-			col[o + 2] = bb
-			col[o + 3] = aa
+			const o = y * 4
+			col2[o] = rr
+			col2[o + 1] = gg
+			col2[o + 2] = bb
+			col2[o + 3] = aa
 		}
-		for (let yy = 0; yy < band; yy++) {
-			const idx = ((y0 + yy) * width + x) * 4
-			const o = yy * 4
-			h2[idx] = col[o]
-			h2[idx + 1] = col[o + 1]
-			h2[idx + 2] = col[o + 2]
-			h2[idx + 3] = col[o + 3]
+		for (let y = 0; y < height; y++) {
+			const idx = (y * width + x) * 4
+			const o = y * 4
+			h2[idx] = col2[o]
+			h2[idx + 1] = col2[o + 1]
+			h2[idx + 2] = col2[o + 2]
+			h2[idx + 3] = col2[o + 3]
 		}
 	}
 
@@ -225,6 +303,7 @@ function gaussianBlurBand(
 
 /**
  * 截图并同时生成两张图: 清晰版 + 竖直条带 [start, end] 高斯模糊版 (签到表脱敏用).
+ * start/end 是「宽度」比例 (从左到右), 条带贯穿全高 (顶边→底边).
  * 同一帧出两张, 避免两次截图不一致.
  */
 export function captureFrameWithBlur(
@@ -246,12 +325,12 @@ export function captureFrameWithBlur(
 	ctx.drawImage(video, 0, 0, w, h)
 	const sharp = canvas.toDataURL('image/jpeg', 0.85)
 
-	const y0 = Math.max(0, Math.floor(h * start))
-	const y1 = Math.min(h, Math.floor(h * end))
+	const x0 = Math.max(0, Math.floor(w * start))
+	const x1 = Math.min(w, Math.floor(w * end))
 	let blurred = sharp
-	if (y1 - y0 > 1) {
+	if (x1 - x0 > 1) {
 		const imgData = ctx.getImageData(0, 0, w, h)
-		const dst = gaussianBlurBand(imgData.data, w, h, y0, y1, radius)
+		const dst = gaussianBlurBand(imgData.data, w, h, x0, x1, radius)
 		imgData.data.set(dst)
 		ctx.putImageData(imgData, 0, 0)
 		blurred = canvas.toDataURL('image/jpeg', 0.85)
diff --git a/ry-vue3/src/layout/AdminLayout.vue b/ry-vue3/src/layout/AdminLayout.vue
index 87cc17d..e5dab4f 100644
--- a/ry-vue3/src/layout/AdminLayout.vue
+++ b/ry-vue3/src/layout/AdminLayout.vue
@@ -195,24 +195,25 @@ const MENU = {
   ],
   doctor: [
     { path: '/doctor/home', title: '首页', icon: House },
-    // 以下 3 项仅审核通过 (audit_status='2') 才显示, 由 menu.filter 用 requireAuditApproved 过滤
+    { path: '/doctor/submissions', title: '我的项目策划方案', icon: EditPen },
+    // 以下 2 项仅审核通过 (audit_status='2') 才显示, 由 menu.filter 用 requireAuditApproved 过滤
     { path: '/doctor/meetings', title: '我参与的会议', icon: Calendar, requireAuditApproved: true },
     { path: '/doctor/projects', title: '我报名的项目', icon: Document, requireAuditApproved: true },
     { path: '/doctor/messages', title: '消息通知', icon: Bell },
-    { path: '/doctor/submissions', title: '我的项目设计投稿', icon: EditPen, requireAuditApproved: true },
     { path: '/doctor/account', title: '账号信息', icon: User }
   ],
   executor: [
     { path: '/executor/overview', title: '首页', icon: House },
+    { path: '/executor/submissions', title: '我的项目策划方案', icon: EditPen },
     { path: '/executor/meetings', title: '会议执行', icon: Calendar },
     { path: '/executor/projects', title: '项目列表', icon: Document },
     { path: '/executor/people', title: '人员管理', icon: User, requireMain: true },
-    { path: '/executor/labor', title: '劳务凭证', icon: EditPen },
     { path: '/executor/messages', title: '消息通知', icon: Bell },
     { path: '/executor/account', title: '账号信息', icon: Setting }
   ],
   sponsor: [
     { path: '/sponsor/home', title: '首页', icon: House },
+    { path: '/sponsor/submissions', title: '我的项目策划方案', icon: EditPen },
     { path: '/sponsor/my-projects', title: '我的项目', icon: Document },
     { path: '/sponsor/meetings', title: '会议列表', icon: Calendar },
     { path: '/sponsor/people', title: '人员管理', icon: User, requireMain: true },
diff --git a/ry-vue3/src/router/index.js b/ry-vue3/src/router/index.js
index 99c8a08..6ec79de 100644
--- a/ry-vue3/src/router/index.js
+++ b/ry-vue3/src/router/index.js
@@ -95,10 +95,10 @@ const routes = [
       { path: 'meetings', name: 'doctor-meetings', component: () => import('@/views/doctor/Meetings.vue'), meta: { title: '我参与的会议' } },
       { path: 'projects', name: 'doctor-projects', component: () => import('@/views/doctor/Projects.vue'), meta: { title: '我报名的项目' } },
       { path: 'messages', name: 'doctor-messages', component: () => import('@/views/messages/Messages.vue'), meta: { title: '消息通知' } },
-      { path: 'submissions', name: 'doctor-submissions', component: () => import('@/views/doctor/Submissions.vue'), meta: { title: '我的项目设计投稿' } },
-      { path: 'submission/new', name: 'doctor-submission-new', component: () => import('@/views/doctor/SubmissionNew.vue'), meta: { title: '新建项目设计投稿' } },
-      { path: 'submission/detail/:planId', name: 'doctor-submission-detail', component: () => import('@/views/doctor/SubmissionDetail.vue'), meta: { title: '投稿详情' } },
-      { path: 'submission/edit/:planId', name: 'doctor-submission-edit', component: () => import('@/views/doctor/SubmissionNew.vue'), meta: { title: '修改项目设计投稿' } },
+      { path: 'submissions', name: 'doctor-submissions', component: () => import('@/views/doctor/Submissions.vue'), meta: { title: '我的项目策划方案' } },
+      { path: 'submission/new', name: 'doctor-submission-new', component: () => import('@/views/doctor/SubmissionNew.vue'), meta: { title: '新建项目策划方案' } },
+      { path: 'submission/detail/:planId', name: 'doctor-submission-detail', component: () => import('@/views/doctor/SubmissionDetail.vue'), meta: { title: '策划方案详情' } },
+      { path: 'submission/edit/:planId', name: 'doctor-submission-edit', component: () => import('@/views/doctor/SubmissionNew.vue'), meta: { title: '修改项目策划方案' } },
       { path: 'account', name: 'doctor-account', component: () => import('@/views/doctor/Account.vue'), meta: { title: '账号信息' } },
       { path: 'sign-fill', name: 'doctor-sign-fill', component: () => import('@/views/doctor/SignFill.vue'), meta: { title: '填写劳务信息', role: 'doctor', hideMenu: true } },
       { path: 'sign-contract', name: 'doctor-sign-contract', component: () => import('@/views/doctor/SignContract.vue'), meta: { title: '签署劳务协议', role: 'doctor', hideMenu: true } },
@@ -108,6 +108,10 @@ const routes = [
   { path: '/executor', component: AdminLayout, meta: { role: 'executor' }, children: [
       { path: '', redirect: { name: 'executor-overview' } },
       { path: 'overview', name: 'executor-overview', component: () => import('@/views/executor/Overview.vue'), meta: { title: '首页' } },
+      { path: 'submissions', name: 'executor-submissions', component: () => import('@/views/doctor/Submissions.vue'), meta: { title: '我的项目策划方案' } },
+      { path: 'submission/new', name: 'executor-submission-new', component: () => import('@/views/doctor/SubmissionNew.vue'), meta: { title: '新建项目策划方案' } },
+      { path: 'submission/detail/:planId', name: 'executor-submission-detail', component: () => import('@/views/doctor/SubmissionDetail.vue'), meta: { title: '策划方案详情' } },
+      { path: 'submission/edit/:planId', name: 'executor-submission-edit', component: () => import('@/views/doctor/SubmissionNew.vue'), meta: { title: '修改项目策划方案' } },
       { path: 'meetings', name: 'executor-meetings', component: () => import('@/views/executor/Meetings.vue'), meta: { title: '会议执行' } },
       { path: 'meetings/new', name: 'executor-meetings-new', component: () => import('@/views/meetings/MeetingNew.vue'), meta: { title: '新建会议' } },
       { path: 'meetings/detail/:meetingId', name: 'executor-meetings-detail', component: () => import('@/views/meetings/MeetingDetail.vue'), meta: { title: '会议详情' } },
@@ -125,6 +129,10 @@ const routes = [
   { path: '/sponsor', component: AdminLayout, meta: { role: 'sponsor' }, children: [
       { path: '', redirect: { name: 'sponsor-home' } },
       { path: 'home', name: 'sponsor-home', component: () => import('@/views/sponsor/Home.vue'), meta: { title: '首页' } },
+      { path: 'submissions', name: 'sponsor-submissions', component: () => import('@/views/doctor/Submissions.vue'), meta: { title: '我的项目策划方案' } },
+      { path: 'submission/new', name: 'sponsor-submission-new', component: () => import('@/views/doctor/SubmissionNew.vue'), meta: { title: '新建项目策划方案' } },
+      { path: 'submission/detail/:planId', name: 'sponsor-submission-detail', component: () => import('@/views/doctor/SubmissionDetail.vue'), meta: { title: '策划方案详情' } },
+      { path: 'submission/edit/:planId', name: 'sponsor-submission-edit', component: () => import('@/views/doctor/SubmissionNew.vue'), meta: { title: '修改项目策划方案' } },
       { path: 'projects', name: 'sponsor-projects', component: () => import('@/views/sponsor/Projects.vue'), meta: { title: '项目管理' } },
       { path: 'my-projects', name: 'sponsor-my-projects', component: () => import('@/views/sponsor/SponsorProjects.vue'), meta: { title: '我的项目' } },
       { path: 'projects/detail/:projectId', name: 'sponsor-projects-detail', component: () => import('@/views/manager/ManagerProjectDetail.vue'), meta: { title: '项目详情' } },
diff --git a/ry-vue3/src/views/doctor/SubmissionDetail.vue b/ry-vue3/src/views/doctor/SubmissionDetail.vue
index 1f2426f..21f4b95 100644
--- a/ry-vue3/src/views/doctor/SubmissionDetail.vue
+++ b/ry-vue3/src/views/doctor/SubmissionDetail.vue
@@ -1,7 +1,7 @@
 <template>
   <div v-loading="loading" class="page-card doctor-submission-detail">
     <!-- 面包屑 (标准: 首页 / 分类 / 当前页) -->
-    <div class="breadcrumb">首页 / 我的项目设计投稿 / 投稿详情</div>
+    <div class="breadcrumb">首页 / 我的项目策划方案 / 策划方案详情</div>
 
     <!-- 表单 (只读, 用 el-input readonly 展示) -->
     <el-form label-width="120px" class="readonly-form">
@@ -68,6 +68,8 @@ import request from '@/utils/request'
 
 const route = useRoute()
 const router = useRouter()
+// 投稿模块路径前缀: 按当前角色 (doctor/executor/sponsor) 拼, 避免串到别的角色侧栏
+const base = computed(() => `/${route.meta.role || 'doctor'}`)
 
 const loading = ref(false)
 const detail = ref({})
@@ -87,7 +89,7 @@ const display = computed(() => ({
 }))
 
 function goBack() {
-  router.push({ path: '/doctor/submissions', query: { _t: Date.now() } })
+  router.push({ path: `${base.value}/submissions`, query: { _t: Date.now() } })
 }
 
 async function load() {
diff --git a/ry-vue3/src/views/doctor/SubmissionNew.vue b/ry-vue3/src/views/doctor/SubmissionNew.vue
index e7aad03..f799c51 100644
--- a/ry-vue3/src/views/doctor/SubmissionNew.vue
+++ b/ry-vue3/src/views/doctor/SubmissionNew.vue
@@ -1,7 +1,7 @@
 <template>
   <div v-loading="loadingDetail" class="page-card doctor-submission-new">
     <div class="breadcrumb">
-      <span>首页 / 我的项目设计投稿 / {{ isEdit ? '修改项目设计投稿' : '新建项目设计投稿' }}</span>
+      <span>首页 / 我的项目策划方案 / {{ isEdit ? '修改项目策划方案' : '新建项目策划方案' }}</span>
     </div>
 
     <el-form :model="form" :rules="rules" ref="formRef" label-width="120px" class="submission-form">
@@ -41,20 +41,14 @@
       <el-row :gutter="12">
         <el-col :span="24">
           <el-form-item label="设计文件" prop="designFileUrl">
-            <el-upload
-              drag
-              action="#"
-              accept=".pdf,.png,.jpg,.jpeg"
-              :show-file-list="true"
-              :file-list="designFileList"
-              :before-upload="beforeDesignUpload"
-              :http-request="uploadDesign"
-              :on-remove="onDesignRemove"
-            >
-              <i class="el-icon-upload"></i>
-              <div class="el-upload__text">将设计文件拖到此处,或<em>点击上传</em></div>
-              <div class="el-upload__tip" slot="tip">支持 PDF / 图片, 单文件 ≤ 20MB</div>
-            </el-upload>
+            <OssFileUploader
+              v-model="form.designFileUrl"
+              dir="ry8080/submission/design/"
+              placeholder="点击上传设计文件"
+              hint="支持 PDF / 图片, 单文件 ≤ 20MB"
+              :max-size="20"
+              block
+            />
           </el-form-item>
         </el-col>
       </el-row>
@@ -76,24 +70,24 @@
 </template>
 
 <script setup>
-import { reactive, ref, onMounted } from 'vue'
+import { reactive, ref, onMounted, computed } from 'vue'
 import { ElMessage, ElMessageBox } from 'element-plus'
 import { useRoute, useRouter } from 'vue-router'
 import { useUserStore } from '@/store/user'
 import request from '@/utils/request'
 import { bizAdd, bizUpdate } from '@/api/public'
-import { uploadToOss } from '@/utils/oss'
 import DictSelect from '@/components/DictSelect.vue'
+import OssFileUploader from '@/components/OssFileUploader.vue'
 
 const router = useRouter()
 const route = useRoute()
 const userStore = useUserStore()
+// 投稿模块路径前缀: 按当前角色 (doctor/executor/sponsor) 拼, 避免串到别的角色侧栏
+const base = computed(() => `/${route.meta.role || 'doctor'}`)
 
 const formRef = ref(null)
 const saving = ref(false)
 const loadingDetail = ref(false)
-const designFileList = ref([])
-const designUploading = ref(false)
 const isEdit = !!route.params.planId
 
 // 项目方向 options (复用 manager 侧接口)
@@ -124,36 +118,6 @@ const rules = {
   designFileUrl: [{ required: true, message: '请上传设计文件', trigger: 'change' }]
 }
 
-function beforeDesignUpload(file) {
-  const max = 20 * 1024 * 1024
-  if (file.size > max) { ElMessage.error('设计文件不能超过 20MB'); return false }
-  const accept = ['.pdf', '.png', '.jpg', '.jpeg']
-  const ext = '.' + (file.name.split('.').pop() || '').toLowerCase()
-  if (!accept.includes(ext)) { ElMessage.error('仅支持 PDF / 图片格式'); return false }
-  return true
-}
-
-async function uploadDesign(opts) {
-  const file = opts.file
-  designUploading.value = true
-  try {
-    const url = await uploadToOss(file, 'ry8080/submission/design/')
-    form.designFileUrl = url
-    designFileList.value = [{ name: file.name, url }]
-    ElMessage.success('设计文件上传成功')
-  } catch (e) {
-    ElMessage.error('上传失败: ' + (e?.message || e))
-    designFileList.value = []
-  } finally {
-    designUploading.value = false
-  }
-}
-
-function onDesignRemove() {
-  form.designFileUrl = ''
-  designFileList.value = []
-}
-
 async function loadDetail() {
   const planId = route.params.planId
   if (!planId) return
@@ -170,10 +134,6 @@ async function loadDetail() {
       form.designFileUrl = d.designFileUrl || ''
       form.status = d.status && String(d.status).trim() ? String(d.status) : '0'
       form.remark = d.remark || ''
-      if (d.designFileUrl) {
-        const name = d.designFileUrl.split('/').pop() || `${d.planName || '设计文件'}.pdf`
-        designFileList.value = [{ name, url: d.designFileUrl }]
-      }
     } else {
       ElMessage.error(res?.msg || '加载失败')
       goBack()
@@ -188,7 +148,7 @@ async function loadDetail() {
 }
 
 function goBack() {
-  router.push({ path: '/doctor/submissions', query: { _t: Date.now() } })
+  router.push({ path: `${base.value}/submissions`, query: { _t: Date.now() } })
 }
 
 function confirmCancel() {
diff --git a/ry-vue3/src/views/doctor/Submissions.vue b/ry-vue3/src/views/doctor/Submissions.vue
index 3b6fe00..cd0175d 100644
--- a/ry-vue3/src/views/doctor/Submissions.vue
+++ b/ry-vue3/src/views/doctor/Submissions.vue
@@ -1,6 +1,6 @@
 <template>
   <div class="page-card">
-    <div class="breadcrumb">首页 / 我的项目设计投稿</div>
+    <div class="breadcrumb">首页 / 我的项目策划方案</div>
 
     <el-form inline :model="q" class="filter-form">
       <el-form-item label="策划方案名称"><el-input v-model="q.planName" placeholder="输入策划方案名称" clearable style="width: 200px" /></el-form-item>
@@ -81,15 +81,18 @@
 </template>
 
 <script setup>
-import { reactive, ref, onMounted } from 'vue'
+import { reactive, ref, onMounted, computed } from 'vue'
 import { ElMessage, ElMessageBox } from 'element-plus'
-import { useRouter } from 'vue-router'
+import { useRouter, useRoute } from 'vue-router'
 import request from '@/utils/request'
 import { bizList, bizUpdate } from '@/api/public'
 import DictSelect from '@/components/DictSelect.vue'
 import AuditStatusTag from '@/components/AuditStatusTag.vue'
 
 const router = useRouter()
+const route = useRoute()
+// 投稿模块路径前缀: 按当前角色 (doctor/executor/sponsor) 拼, 避免串到别的角色侧栏
+const base = computed(() => `/${route.meta.role || 'doctor'}`)
 
 const q = reactive({
   planName: '',
@@ -146,13 +149,13 @@ function canSubmit(row) {
 function onSelectionChange(arr) { selected.value = arr }
 
 function onCreate() {
-  router.push({ path: '/doctor/submission/new' })
+  router.push({ path: `${base.value}/submission/new` })
 }
 function onView(row) {
-  router.push({ path: `/doctor/submission/detail/${row.planId}` })
+  router.push({ path: `${base.value}/submission/detail/${row.planId}` })
 }
 function onEdit(row) {
-  router.push({ path: `/doctor/submission/edit/${row.planId}` })
+  router.push({ path: `${base.value}/submission/edit/${row.planId}` })
 }
 
 async function onSubmit(row) {
diff --git a/ry-vue3/src/views/portal/Home.vue b/ry-vue3/src/views/portal/Home.vue
index c83c679..48eb90c 100644
--- a/ry-vue3/src/views/portal/Home.vue
+++ b/ry-vue3/src/views/portal/Home.vue
@@ -48,7 +48,7 @@
           </p>
         </div>
         <div class="hero-cta">
-          <span class="cta-btn" @click="onSubmit">项目提案</span>
+          <span v-if="canPropose" class="cta-btn" @click="onSubmit">项目提案</span>
         </div>
       </section>
 
@@ -169,6 +169,11 @@ function openPlan(id) {
 const topNavClass = computed(() => isScrolled.value ? 'is-scrolled' : '')
 const loggedIn = computed(() => !!userStore.token)
 const userName = computed(() => userStore.user?.userName || '用户')
+// 投稿角色: admin/manager 不开放投稿 → 隐藏「项目提案」按钮; 未登录 user 为 null → 显示(点击引导登录)
+const canPropose = computed(() => {
+  const r = userStore.user?.role || ''
+  return r !== 'admin' && r !== 'manager'
+})
 
 const planSvg = `<svg viewBox="0 150 1200 370" xmlns="http://www.w3.org/2000/svg" role="img" aria-label="2025-2030年项目规划">
 <defs>
@@ -257,7 +262,9 @@ function onSubmit() {
     router.push('/login')
     return
   }
-  router.push('/publicity')
+  const role = userStore.user?.role
+  const map = { doctor: '/doctor/submissions', executor: '/executor/submissions', sponsor: '/sponsor/submissions' }
+  router.push(map[role] || '/doctor/submissions')
 }
 </script>