"""从 OCR 文本/行里抽取发票字段 适配中国大陆 增值税发票（电子普票 / 专票 / 电子专票 / 数电票） 关键策略： - 主体（销售方/购买方）按 OCR box **坐标**判断归属（左右两栏） - 名称提取加 stop word，避免 OCR 单行文本混淆 """ from __future__ import annotations import re from typing import List, Optional, Tuple from app.models import InvoiceFields, OCRLine from app.utils import ( amount_consistent, extract_cn_amount, extract_num_amount, extract_pretax_amount, extract_tax_amount, extract_total_amount, ) # ---------- 发票类型 ---------- _INVOICE_TYPES = [ "增值税电子专用发票", "增值税电子普通发票", "增值税专用发票", "增值税普通发票", "通用机打发票", "数电票", "电子发票", ] # ---------- 发票号码 ---------- _NO_PATTERN = re.compile( r"(?:发\s*票\s*号\s*码|号\s*码|No\.?|号)\s*[:：]?\s*(\d{8,20})", re.IGNORECASE, ) # ---------- 发票代码 ---------- _CODE_PATTERN = re.compile( r"(?:发\s*票\s*代\s*码|代\s*码)\s*[:：]?\s*(\d{10,12}|\d{8,12})", ) # ---------- 开票日期 ---------- _DATE_PATTERN = re.compile( r"(?:开\s*票\s*日\s*期|日\s*期)\s*[:：]?\s*" r"(\d{4})\s*[年/\.]\s*(\d{1,2})\s*[月/\.]\s*(\d{1,2})", ) # ---------- 纳税人识别号（必须至少含 1 个字母，排除纯数字发票号）---------- _TAX_NO_PATTERN = re.compile(r"((?=[0-9A-Z]*[A-Z])[0-9A-Z]{18})") # ---------- 主体标签 ---------- _BUYER_LABEL = re.compile(r"购\s*买\s*方\s*(?:信\s*息|名\s*称|)") _SELLER_LABEL = re.compile(r"销\s*售\s*方\s*(?:信\s*息|名\s*称|)") # ---------- 名称（带 stop word 截断）---------- _NAME_STOP = r"(?: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)" _NAME_PATTERN = re.compile( rf"名\s*称\s*[:：]\s*" rf"((?:(?!{_NAME_STOP})[^\n\r]){{2,60}}?(?:公司|商店|厂|店|部|中心|工作室))" ) # ---------- 工具函数 ---------- def _norm(text: str) -> str: return re.sub(r"\s+", " ", text or "").strip() def _box_center(box) -> Tuple[float, float]: """box: [[x1,y1], [x2,y2], [x3,y3], [x4,y4]] → (cx, cy)""" if not box or len(box) < 4: return (0.0, 0.0) xs = [p[0] for p in box] ys = [p[1] for p in box] return ((min(xs) + max(xs)) / 2, (min(ys) + max(ys)) / 2) def _detect_invoice_type(text: str) -> Optional[str]: for t in _INVOICE_TYPES: if t in text: return t return None def _extract_invoice_no(text: str) -> Optional[str]: m = _NO_PATTERN.search(text) return m.group(1) if m else None def _extract_invoice_code(text: str) -> Optional[str]: m = _CODE_PATTERN.search(text) return m.group(1) if m else None def _extract_date(text: str) -> Optional[str]: m = _DATE_PATTERN.search(text) if not m: return None y, mo, d = m.groups() return f"{int(y):04d}-{int(mo):02d}-{int(d):02d}" def _clean_name(name: str) -> str: name = re.sub(r"^[\s:：,，。、]+", "", name) name = name.split("纳税人")[0].split("统一社会")[0] name = re.sub(r"[^一-龥A-Za-z0-9（）()·\-]", "", name) return name.rstrip("：:;,，。、 ").strip() def _extract_parties_from_lines(lines: List[OCRLine]) -> tuple: """基于 box 坐标的主体识别 1) 找到 "购买方信息"/"销售方信息" 标签所在位置（box 中心） 2) 把每个 "名称: xxx" 行按 x 坐标归属到对应栏 3) 兜底：左右栏无法区分时按 y 坐标（top/bottom） """ # 找标签坐标 buyer_label_box = None seller_label_box = None for line in lines: if _BUYER_LABEL.search(line.text) and not buyer_label_box: buyer_label_box = line.box if _SELLER_LABEL.search(line.text) and not seller_label_box: seller_label_box = line.box by_x = buyer_label_box is not None and seller_label_box is not None and \ abs(_box_center(buyer_label_box)[0] - _box_center(seller_label_box)[0]) > 50 # 找所有 "名称: ..." 行的 box name_lines = [] for line in lines: m = _NAME_PATTERN.search(line.text) if m: name_lines.append((line, _clean_name(m.group(1)))) # 找所有税号 tax_lines = [] for line in lines: m = _TAX_NO_PATTERN.search(line.text) if m: tax_lines.append((line, m.group(1))) seller_name = buyer_name = None seller_tax = buyer_tax = None if by_x and buyer_label_box and seller_label_box: # 左右栏布局：用 x 坐标归属 bx, _ = _box_center(buyer_label_box) sx, _ = _box_center(seller_label_box) mid = (bx + sx) / 2 for line, name in name_lines: cx, _ = _box_center(line.box) if cx < mid and not buyer_name: buyer_name = name elif cx >= mid and not seller_name: seller_name = name for line, tax in tax_lines: cx, _ = _box_center(line.box) if cx < mid and not buyer_tax: buyer_tax = tax elif cx >= mid and not seller_tax: seller_tax = tax else: # 上下栏 / 单栏：按 y 坐标（第一个 = 购买方） name_lines.sort(key=lambda x: (_box_center(x[0].box)[1], _box_center(x[0].box)[0])) tax_lines.sort(key=lambda x: (_box_center(x[0].box)[1], _box_center(x[0].box)[0])) if len(name_lines) >= 1: buyer_name = name_lines[0][1] if len(name_lines) >= 2: seller_name = name_lines[1][1] if len(tax_lines) >= 1: buyer_tax = tax_lines[0][1] if len(tax_lines) >= 2: seller_tax = tax_lines[1][1] return seller_name, seller_tax, buyer_name, buyer_tax def _extract_parties_from_text(text: str) -> tuple: """无 box 信息时的 fallback：按显式标签分段""" seller_name = buyer_name = None seller_tax = buyer_tax = None for m in _BUYER_LABEL.finditer(text): chunk = text[m.end():m.end() + 200] nm = _NAME_PATTERN.search(chunk) tm = _TAX_NO_PATTERN.search(chunk[:100]) if nm and not buyer_name: buyer_name = _clean_name(nm.group(1)) if tm and not buyer_tax: buyer_tax = tm.group(1) if buyer_name and buyer_tax: break for m in _SELLER_LABEL.finditer(text): chunk = text[m.end():m.end() + 200] nm = _NAME_PATTERN.search(chunk) tm = _TAX_NO_PATTERN.search(chunk[:100]) if nm and not seller_name: seller_name = _clean_name(nm.group(1)) if tm and not seller_tax: seller_tax = tm.group(1) if seller_name and seller_tax: break if not buyer_name: name_positions = [(m.start(), m.group(1)) for m in _NAME_PATTERN.finditer(text)] if len(name_positions) >= 1: buyer_name = _clean_name(name_positions[0][1]) if len(name_positions) >= 2: seller_name = _clean_name(name_positions[1][1]) tax_positions = _TAX_NO_PATTERN.findall(text) if not buyer_tax and len(tax_positions) >= 1: buyer_tax = tax_positions[0] if not seller_tax and len(tax_positions) >= 2: seller_tax = tax_positions[1] return seller_name, seller_tax, buyer_name, buyer_tax def extract(text: str, lines: Optional[List[OCRLine]] = None) -> InvoiceFields: raw = _norm(text) fields = InvoiceFields() fields.invoice_type = _detect_invoice_type(raw) fields.invoice_code = _extract_invoice_code(raw) fields.invoice_no = _extract_invoice_no(raw) fields.invoice_date = _extract_date(raw) # 金额 total = extract_total_amount(raw) tax = extract_tax_amount(raw) pretax = extract_pretax_amount(raw) # 兜底：从所有 .xx 数字中找满足 tax + pretax = total 的组合 if (tax is None or pretax is None) and total is not None: candidates = sorted( {float(x) for x in re.findall(r"(\d+\.\d{2})", raw) if float(x) < total}, reverse=True, ) for i, a in enumerate(candidates): for b in candidates[i + 1:]: if abs(a + b - total) < 0.011: if pretax is None: pretax = round(a, 2) if tax is None: tax = round(b, 2) break if tax is not None and pretax is not None: break # 单数字兜底：如果 total 不在候选里，只剩一个候选时 if (tax is None or pretax is None) and len(candidates) == 1: only = round(candidates[0], 2) if pretax is None and tax is None: pretax = only tax = round(total - only, 2) elif tax is None: tax = only elif pretax is None: pretax = only # 二次兜底：合计 - 任一 = 另一 if tax is None and total is not None and pretax is not None: tax = round(total - pretax, 2) if pretax is None and total is not None and tax is not None: pretax = round(total - tax, 2) fields.amount = total fields.tax_amount = tax fields.amount_pretax = pretax fields.amount_cn = extract_cn_amount(raw) fields.amount_match = amount_consistent(fields.amount_cn, fields.amount) # 主体 if lines: seller_name, seller_tax, buyer_name, buyer_tax = _extract_parties_from_lines(lines) if not (seller_name and buyer_name): sn, st, bn, bt = _extract_parties_from_text(raw) seller_name = seller_name or sn buyer_name = buyer_name or bn seller_tax = seller_tax or st buyer_tax = buyer_tax or bt else: seller_name, seller_tax, buyer_name, buyer_tax = _extract_parties_from_text(raw) fields.seller_name = seller_name fields.seller_tax_no = seller_tax fields.buyer_name = buyer_name fields.buyer_tax_no = buyer_tax return fields