郭庆泰
c3eb8ed9c3
ry-ocr/ (新)
本地发票识别微服务 (PaddleOCR 3.x + FastAPI, 8801)
- QR 优先: 扫到二维码即取开票时间/发票号/金额; 没扫到/格式不合法直接判非发票, 不跑 OCR
- 配置 QR_FULL_OCR 控制快路径(false, 0.2s)还是全字段(true, 4.5s)
- /recognize/invoice (multipart) + /recognize/invoice/by-path (本地路径, 白名单) + /recognize/text
- is_invoice / from_qr / qr_raw / qr_error / error_code 字段
- 12 字段发票抽取 (regex + 启发式, 左右主体识别)
- 超时保护 (15s 单页 / 60s 总流程) + PaddleOCR 单例 + ThreadPoolExecutor
ry-api/ruoyi-business/
- pom.xml: 加 hutool-http/json/core 5.8.27, lombok 1.18.30 (OcrClient @Slf4j 所需)
- ocr/: OcrClient + InvoiceResult/Fields/Line + ZipExtractor + InvoiceOcrScheduler
- oss/: OssUploader + OssConfMeta (OCR 识别后重传 OSS)
- config/: OcrConfig + OcrExecutorConfig (后台线程池)
- service/impl/InvoiceOcrService: 后台提交 OCR, ZIP 路径解压识别, 替换场景先清旧
- 会议材料 CRUD 全套 (BizMeetingAuditLog/Executor/Invoice/Material/Supervisor):
controller + service + mapper + domain + xml
ry-vue3/
- MeetingDetail.vue (新建): 会议详情页 (含评分维度章节, 改只读)
- Meetings.vue / OssFileUploader.vue / router / Login.vue: 适配新字段
ry-api/ruoyi-admin/
- RuoYiApplication.java + application.yml: 启用 @Async 异步支持
_self/
- manager_meetings.md / manager_meeting_detail.md: 文档
93 lines
2.9 KiB
Python
93 lines
2.9 KiB
Python
"""基准测试：不同 DPI 下的 PDF 渲染耗时 vs OCR 耗时 vs 识别准确度"""
|
|
import sys
|
|
import time
|
|
from pathlib import Path
|
|
|
|
try:
|
|
sys.stdout.reconfigure(encoding="utf-8")
|
|
except Exception:
|
|
pass
|
|
|
|
ROOT = Path(__file__).resolve().parent.parent
|
|
sys.path.insert(0, str(ROOT))
|
|
|
|
import fitz
|
|
from app.core import warmup
|
|
from app.services.recognize_service import recognize_file
|
|
|
|
|
|
def render(pdf_path: Path, dpi: int) -> Path:
|
|
out = ROOT / f"_bench_{dpi}dpi.png"
|
|
doc = fitz.open(str(pdf_path))
|
|
page = doc[0]
|
|
pix = page.get_pixmap(matrix=fitz.Matrix(dpi / 72, dpi / 72), alpha=False)
|
|
pix.save(str(out))
|
|
doc.close()
|
|
return out
|
|
|
|
|
|
def bench_dpi(pdf_path: Path, dpi: int, runs: int = 3):
|
|
img_path = render(pdf_path, dpi)
|
|
img_bytes = img_path.read_bytes()
|
|
print(f"\n--- DPI={dpi} | 渲染图 {img_path.stat().st_size//1024}KB ---")
|
|
|
|
times = []
|
|
fields_ok = None
|
|
for i in range(runs):
|
|
t0 = time.time()
|
|
result = recognize_file(img_path.name, img_bytes)
|
|
elapsed = (time.time() - t0) * 1000
|
|
times.append(elapsed)
|
|
if i == 0 and result.success:
|
|
fields_ok = result.fields
|
|
print(f" run {i+1}: {elapsed:7.1f}ms success={result.success}")
|
|
|
|
# 清理中间文件
|
|
img_path.unlink(missing_ok=True)
|
|
(ROOT / f".{img_path.stem}_pages").rmdir() if (ROOT / f".{img_path.stem}_pages").exists() else None
|
|
# 上面可能删不了子目录，try 再清理 _pages 下文件
|
|
pages_dir = ROOT / f".{img_path.stem}_pages"
|
|
if pages_dir.exists():
|
|
for f in pages_dir.glob("*"):
|
|
f.unlink(missing_ok=True)
|
|
pages_dir.rmdir()
|
|
|
|
times.sort()
|
|
print(f" min={times[0]:.1f}ms median={times[len(times)//2]:.1f}ms max={times[-1]:.1f}ms")
|
|
|
|
return fields_ok
|
|
|
|
|
|
if __name__ == "__main__":
|
|
pdf = ROOT / "fapiao.pdf"
|
|
print(f"PDF: {pdf.name}, {pdf.stat().st_size//1024}KB, 1 页")
|
|
|
|
warmup()
|
|
print("预热完成\n")
|
|
|
|
print("=" * 60)
|
|
fields_100 = bench_dpi(pdf, 100)
|
|
print("=" * 60)
|
|
fields_150 = bench_dpi(pdf, 150)
|
|
print("=" * 60)
|
|
fields_200 = bench_dpi(pdf, 200)
|
|
print("=" * 60)
|
|
fields_300 = bench_dpi(pdf, 300)
|
|
print("=" * 60)
|
|
|
|
# 汇总
|
|
print("\n\n=== 关键字段对比 ===")
|
|
print(f"{'字段':<15} {'100dpi':<25} {'150dpi':<25} {'200dpi':<25} {'300dpi':<25}")
|
|
key_fields = ["invoice_no", "invoice_date", "amount", "amount_cn",
|
|
"amount_pretax", "tax_amount", "seller_name", "buyer_name",
|
|
"seller_tax_no", "buyer_tax_no", "amount_match"]
|
|
for k in key_fields:
|
|
row = [k]
|
|
for f in [fields_100, fields_150, fields_200, fields_300]:
|
|
v = getattr(f, k) if f else None
|
|
if v is None:
|
|
row.append("(空)")
|
|
else:
|
|
s = str(v)[:23]
|
|
row.append(s)
|
|
print(f"{row[0]:<15} {row[1]:<25} {row[2]:<25} {row[3]:<25} {row[4]:<25}") |