# -*- coding: utf-8 -*- """ OCR 准确率核对工具 — 用截图当标准答案，核对 Excel 采集数据 ============================================================================= 用法: python audit_ocr.py # 默认核对 采集数据/ 最新一天的 xlsx python audit_ocr.py 采集数据/20260909.xlsx # 指定文件 python audit_ocr.py xxx.xlsx --max 10 # 最多核对10张截图（PP-OCR较慢，约30~60秒/张） 核对口径: A. 店名字准率: Excel里 列表页OCR店名 vs 详情页AI店名（独立来源互为答案） B. 现场覆盖率: 来源截图列 追溯到原图 → 本地PP-OCR重新识别 → 图上枚举出的完整卡片 是否都进了 Excel（漏采检测）；Excel行记录的价格是否真的出现在原图上（追溯检测） 说明: B 口径的图上枚举用本地PP-OCR+几何规则（价格块+下方店铺块+上方标题块），与GLM无关， 所以能发现 GLM 漏卡/错配；字级别的标题准确率仍以人工抽查为准。 """ import json import random import re import sys from pathlib import Path ROOT = Path(__file__).parent sys.path.insert(0, str(ROOT)) MAX_GAP = 216 # 价格→店铺窗口（与采集端校验一致，按1220x2712标定） TITLE_WIN = 330 # 标题窗口（价格上方） PROMO_RE = re.compile(r"起送|满\d|配送|发货|日达|红包|已售|月售|预估|减\d|处方|OTC|同款|低价|原研|新客|超值") def norm(s): return re.sub(r"\s+", "", str(s or "")).lower() def cy(b): return (b["box"][1] + b["box"][3]) // 2 def cx(b): return (b["box"][0] + b["box"][2]) // 2 def price_num(p): m = re.search(r"[\d.]+", str(p or "")) return m.group(0).rstrip(".") if m else "" def edit_dist(a, b): if abs(len(a) - len(b)) > 2: return 99 prev = list(range(len(b) + 1)) for i, ca in enumerate(a, 1): cur = [i] for j, cb in enumerate(b, 1): cur.append(min(prev[j] + 1, cur[j - 1] + 1, prev[j - 1] + (ca != cb))) prev = cur return prev[-1] def shop_match(a, b): na, nb = norm(a), norm(b) if len(na) < 3 or len(nb) < 3: return False if na in nb or nb in na: return True return edit_dist(na, nb) <= 1 def load_excel(path): from openpyxl import load_workbook wb = load_workbook(path, read_only=True) rows = [] for ws in wb.worksheets: for r in list(ws.iter_rows(values_only=True))[1:]: t, p, s1, s2, snap, link, src = (list(r) + [None] * 7)[:7] rows.append({"sheet": ws.title, "title": str(t or ""), "price": str(p or ""), "shop": str(s1 or ""), "shop_ai": str(s2 or ""), "src": str(src or "")}) return rows def shop_accuracy(rows): """口径A: 列表OCR店名 vs 详情页AI店名""" comparable = [r for r in rows if norm(r["shop"]) and norm(r["shop_ai"])] exact = sum(1 for r in comparable if norm(r["shop"]) == norm(r["shop_ai"])) fuzzy = [r for r in comparable if norm(r["shop"]) != norm(r["shop_ai"]) and shop_match(r["shop"], r["shop_ai"])] diff = [r for r in comparable if norm(r["shop"]) != norm(r["shop_ai"]) and not shop_match(r["shop"], r["shop_ai"])] print("=" * 78) print(f"口径A 店名字准率（可比 {len(comparable)}/{len(rows)} 行）") print(f" 完全一致 {exact:>4} {exact/len(comparable)*100 if comparable else 0:.1f}%") print(f" 同店近似 {len(fuzzy):>4} （错1~2字/前后缀差）") print(f" 不一致 {len(diff):>4} ← 逐条人工判断是'读错字'还是'配错卡'") for r in diff: print(f" [{r['sheet']}] {r['title'][:30]} {r['price']}") print(f" 列表OCR: {r['shop'][:24]} ↔ 详情AI: {r['shop_ai'][:24]} 来源:{r['src']}") n = len(comparable) if n: print(f" → 店名识别成功率 ≈ {(exact + len(fuzzy)) / n * 100:.1f}%") return diff def split_right_col(blocks): """定位右列分界x（同采集端：￥块左边缘中位数-15，图片区在左、文字区在右）""" price_x = sorted(b["box"][0] for b in blocks if "¥" in b["text"] or "￥" in b["text"]) if len(price_x) < 3: return 0 return max(price_x[len(price_x) // 2] - 15, 0) def enumerate_cards(blocks): """几何规则枚举图上的完整卡片: 价格块 + 下方店铺块 + 上方标题块（与GLM无关的独立读图） 先裁掉左列商品图（包装文字会污染店铺/标题），价格块必须以¥开头（排除'起送¥0'类促销行）""" split = split_right_col(blocks) blocks = [b for b in blocks if cx(b) >= split] if split else blocks cards = [] used_shop = [] prices = [b for b in blocks if norm(b["text"]).startswith(("¥", "￥")) and re.search(r"¥[\d.]+|￥[\d.]+", b["text"])] prices.sort(key=cy) for pb in prices: py = cy(pb) pn = price_num(pb["text"]) if not pn: continue # 店铺: 价格下方最近的有效块 shop = None for b in sorted(blocks, key=cy): y = cy(b) if not (py + 5 < y < py + MAX_GAP): continue t = norm(b["text"]) if len(t) < 4 or "¥" in b["text"] or "￥" in b["text"] or PROMO_RE.search(t): continue if any(abs(y - u) < 25 for u in used_shop): continue shop = b break if shop is None: continue used_shop.append(cy(shop)) # 标题: 价格上方最长有效块 title, tlen = "", 0 for b in blocks: y = cy(b) if not (py - TITLE_WIN < y < py - 5): continue t = norm(b["text"]) if len(t) < 6 or "¥" in b["text"] or "￥" in b["text"] or PROMO_RE.search(t): continue if len(t) > tlen: title, tlen = b["text"], len(t) cards.append({"price": pn, "shop": re.split(r"[(（]", re.sub(r"\s+", "", shop["text"]))[0], "title": re.sub(r"\s+", "", title), "py": py}) return cards def field_audit(rows, max_images): """口径B: 来源截图追溯 + 图上卡片漏采检测""" src_rows = [r for r in rows if r["src"]] if not src_rows: print("\n口径B 跳过：该Excel没有「来源截图」列数据（列是今天新加的，明天采集后生效）") return try: from ai_helper_vision1 import VisionParser1 except Exception as e: print(f"\n口径B 跳过：PP-OCR加载失败({e})") return # 来源 → 截图文件（取 base 文件；若被覆盖可用同名时间戳副本里最新的一张） shots = {} for r in src_rows: dev, batch = r["src"].split("/step3_b", 1) base = ROOT / "screenshots" / dev / "step3" / f"step3_b{batch}.png" if base.exists(): shots.setdefault(str(base), (dev, batch)) else: # 找时间戳副本 cands = sorted((ROOT / "screenshots" / dev / "step3").glob(f"step3_b{batch}_*.png"), key=lambda p: p.stat().st_mtime) if cands: shots.setdefault(str(cands[-1]), (dev, batch)) parser = VisionParser1() keys = list(shots.keys()) if max_images and len(keys) > max_images: keys = random.sample(keys, max_images) # 跨设备随机抽样,避免样本偏斜 total_cards = matched = 0 missing_report, trace_fail = [], [] audited_srcs = {shots[k] for k in keys} print("\n" + "=" * 78) print(f"口径B 现场核对（{len(keys)}/{len(shots)} 张截图，PP-OCR本地识别，每张约30~60秒）") for k in keys: dev, batch = shots[k] try: blocks = parser._ppocr_blocks(k, 0) # 整图识别，由几何规则自行分列 except Exception as e: print(f" [跳过] {dev}/b{batch}: OCR失败 {e}") continue cards = enumerate_cards(blocks) # 图上卡片 → Excel 全量匹配（价格+店名） hit = 0 for c in cards: rows_p = [r for r in rows if price_num(r["price"]) == c["price"]] if any(shop_match(c["shop"], r["shop"]) for r in rows_p): hit += 1 else: missing_report.append((dev, batch, c)) total_cards += len(cards) matched += hit print(f" {dev}/b{batch}: 图上{len(cards)}卡, 入Excel {hit}") # 追溯: 来源=这张图的Excel行,其价格必须出现在图上 for r in src_rows: if (r["src"].split("/step3_b")[0], r["src"].split("/step3_b")[1]) == (dev, batch): pn = price_num(r["price"]) if pn and not any(pn in norm(b["text"]) for b in blocks if "¥" in b["text"] or "￥" in b["text"]): trace_fail.append(r) print("-" * 78) if total_cards: print(f"图上卡片共 {total_cards} 张, 已入Excel {matched} 张 → 现场覆盖率 {matched/total_cards*100:.1f}%") for dev, batch, c in missing_report: print(f" [图上有但Excel没有] {dev}/b{batch} ¥{c['price']} {c['shop'][:14]} {c['title'][:26]}") if trace_fail: print(f"追溯异常 {len(trace_fail)} 行（Excel记的价格在来源截图上找不到）:") for r in trace_fail[:10]: print(f" [{r['sheet']}] {r['title'][:26]} {r['price']} 来源:{r['src']}") if not trace_fail: print("追溯检测: 全部Excel行的价格都能在其来源截图上找到 ✓") def main(): args = [a for a in sys.argv[1:] if not a.startswith("--")] max_images = None if "--max" in sys.argv: try: max_images = int(sys.argv[sys.argv.index("--max") + 1]) except (IndexError, ValueError): pass if args: xlsx = ROOT / args[0] else: cands = sorted((ROOT / "采集数据").glob("*.xlsx"), key=lambda p: p.stat().st_mtime) if not cands: print("采集数据/ 下没有 xlsx"); sys.exit(1) xlsx = cands[-1] if not xlsx.exists(): print(f"文件不存在: {xlsx}"); sys.exit(1) print(f"核对文件: {xlsx}\n") rows = load_excel(xlsx) print(f"共 {len(rows)} 行, {len(set(r['sheet'] for r in rows))} 个药品\n") shop_accuracy(rows) field_audit(rows, max_images) print("\n提示: 口径B依赖「来源截图」列（今天新加），旧数据该列为空会自动跳过。") if __name__ == "__main__": main()