diff --git a/code_map.json b/code_map.json new file mode 100644 index 0000000..75d654e --- /dev/null +++ b/code_map.json @@ -0,0 +1,20 @@ +{ + "2241.01.05": "上料押金", + "1001.01": "人民币", + "1221.01": "员工借款", + "1012.01": "POS刷卡未到账", + "6051.08": "转售水电费收入", + "2203.01": "物业管理费", + "1002.01.01": "淇县支行1710020809200182975", + "6603.02": "手续费", + "6051.02": "办证工本收入", + "1002.02.01": "漓江支行4105016228400000750", + "1411.02": "电子设备", + "6401.51.09": "日常维修费", + "6401.02.02": "办公费", + "6401.01.02": "职工福利费", + "6402.01": "付垃圾清运费", + "2241.01.01": "装修押金", + "6051.01": "垃圾清运费", + "6051.05": "家政服务费" +} \ No newline at end of file diff --git a/config.yaml b/config.yaml index 64eea92..5cf9ef9 100644 --- a/config.yaml +++ b/config.yaml @@ -1,6 +1,6 @@ ocr: det_db_box_thresh: 0.5 - det_db_score_mode: fast + det_db_score_mode: slow det_db_thresh: 0.3 det_db_unclip_ratio: 1.5 lang: ch diff --git a/core/parser.py b/core/parser.py index 84405c4..5a66934 100644 --- a/core/parser.py +++ b/core/parser.py @@ -19,6 +19,75 @@ class K3VoucherParser: def __init__(self, lines): """lines: [(text, conf, cx, cy, box), ...]""" self.lines = sorted(lines, key=lambda l: (l[3], l[2])) + self._merge_continuation_lines() + + def _merge_continuation_lines(self): + """合并续行:会计科目名称被OCR分成多行时,合并回一行 + + 策略:续行以横杠开头(如 -淇县支行1710020809200182975), + 找到上方最近的一个以横杠结尾的代码行进行合并。 + 如果找不到以横杠结尾的代码行,则找上方最近的不以横杠开头的代码行。 + """ + # 先识别所有代码行 + code_lines = [] # [(idx, text, ends_with_dash, cy, cx)] + for i, (text, conf, cx, cy, box) in enumerate(self.lines): + t = text.strip() + if re.match(r'\d{3,4}\.\d', t): + code_lines.append((i, t, t.endswith('-'), cy, cx)) + + if not code_lines: + return + + merged = list(self.lines) + removed = set() + merged_codes = set() # 已接收续行的代码行,不再接收 + + for i, (text, conf, cx, cy, box) in enumerate(self.lines): + t = text.strip() + if not (t.startswith('-') or t.startswith('-') or t.startswith('—')): + continue + if i in removed: + continue + + # 跳过包含日期/表头关键词的续行 + stripped = t.lstrip('--—') + if any(k in stripped for k in ['日期', '记账凭证', '记账', '业务日期', '凭证号', '附件数']): + continue + + # 匹配代码行 + best_code = None + best_score = 99999 + + for idx, ct, ends_dash, ccy, ccx in code_lines: + if idx in merged_codes: + continue # 已合并过的代码行不再接收续行 + if abs(cy - ccy) > 300: + continue + if abs(cx - ccx) > 300: + continue + + dist_y = abs(cy - ccy) + dist_x = abs(cx - ccx) + + if ends_dash: + score = dist_y * 2 + dist_x * 2 + else: + # 不以横杠结尾的代码行:x必须非常接近才接受 + if dist_x > 100: + continue + score = 1500 + dist_y + dist_x * 2 + + if score < best_score: + best_score = score + best_code = idx + + if best_code is not None and best_score < 2000: + orig_text, orig_conf, orig_cx, orig_cy, orig_box = merged[best_code] + merged[best_code] = (orig_text + text, min(orig_conf, conf), orig_cx, orig_cy, orig_box) + removed.add(i) + merged_codes.add(best_code) # 标记已合并 + + self.lines = [merged[i] for i in range(len(merged)) if i not in removed] def parse(self): result = { @@ -185,7 +254,9 @@ class K3VoucherParser: # 提取名称 — 按用户规则:取最后一个横杠之后的文字 name = t[m.end():].lstrip('--—= ') - name = re.sub(r'[\d.\-—=]+$', '', name).strip() + # 保留长数字尾(银行账号),仅清除短数字尾 + if not re.search(r'\d{6,}', name): + name = re.sub(r'[\d.\-—=]+$', '', name).strip() name = name.lstrip('-—-=') if '\\' in name: @@ -199,7 +270,7 @@ class K3VoucherParser: if code_part in code_map: existing = codes[code_map[code_part]] # 已有名称为空,而新行有名称 → 更新 - if (not existing['name'] or len(existing['name']) < 2) and len(name) >= 2: + if len(name) >= 2 and (not existing['name'] or len(existing['name']) < 2 or len(name) > len(existing['name'])): # 优先长名 existing['name'] = name existing['y_pos'] = cy existing['x_pos'] = cx @@ -251,17 +322,19 @@ class K3VoucherParser: for t, cy, cx in name_candidates: if cy <= code['y_pos']: continue # 只找下方的 - if abs(cx - code['x_pos']) > 600: + if abs(cx - code['x_pos']) > 400: continue dist = cy - code['y_pos'] - if dist < best_dist and dist < 600: + if dist < best_dist and dist < 400: best = (t, cy, cx) best_dist = dist if best: t, cy, cx = best # 按用户规则提取名称(原始逻辑:先清洗后取最后横杠) n = t.lstrip('--—= ') - n = re.sub(r'[\d.\-—=]+$', '', n).strip() + # 保留长数字尾(银行账号) + if not re.search(r'\d{6,}', n): + n = re.sub(r'[\d.\-—=]+$', '', n).strip() n = n.lstrip('-—-=') if '\\' in n: n = n.split('\\', 1)[0].strip() @@ -286,17 +359,19 @@ class K3VoucherParser: for nt, ncy, ncx in name_candidates: if ncy <= code['y_pos']: continue - if abs(ncx - code['x_pos']) > 600: + if abs(ncx - code['x_pos']) > 400: continue ndist = ncy - code['y_pos'] - if ndist < next_dist and ndist < 600: + if ndist < next_dist and ndist < 400: next_best = (nt, ncy, ncx) next_dist = ndist if not next_best: break nt, ncy, ncx = next_best nn = nt.lstrip('--—= ') - nn = re.sub(r'[\d.\-—=]+$', '', nn).strip() + # 保留长数字尾(银行账号) + if not re.search(r'\d{6,}', nn): + nn = re.sub(r'[\d.\-—=]+$', '', nn).strip() nn = nn.lstrip('-—-=') if '\\' in nn: nn = nn.split('\\', 1)[0].strip() diff --git a/core/recognizer.py b/core/recognizer.py index 7f1ff08..52bbe80 100644 --- a/core/recognizer.py +++ b/core/recognizer.py @@ -27,18 +27,29 @@ def get_ocr_instance(config: dict): return _ocr -def run_ocr(image, ocr_config: dict) -> list: +def run_ocr(image, ocr_config: dict, photo_type: str = 'normal') -> list: """对图像执行 OCR,返回 [(text, confidence, cx, cy, box), ...] 输入: image: numpy array (cv2.imread) 或 图像文件路径 ocr_config: dict(PaddleOCR 参数) + photo_type: str — 照片类型,用于动态调参 返回: 排序后的 lines: [(text, conf, cx, cy, box), ...] 按 (y, x) 排序 """ - ocr = get_ocr_instance(ocr_config) + # 根据照片类型动态调整 OCR 参数 + config_override = dict(ocr_config) + if photo_type == 'flash': + config_override['det_db_thresh'] = config_override.get('det_db_thresh_flash', 0.18) + elif photo_type == 'dark': + config_override['det_db_thresh'] = config_override.get('det_db_thresh_dark', 0.25) + elif photo_type == 'blurry': + config_override['det_db_thresh'] = config_override.get('det_db_thresh_blurry', 0.2) + # normal 保持默认 + + ocr = get_ocr_instance(config_override) # 支持 numpy array 和 文件路径 result = ocr(image) diff --git a/kocr.py b/kocr.py index 2a2fae9..31bcf05 100644 --- a/kocr.py +++ b/kocr.py @@ -22,6 +22,7 @@ KOCR v4 — 金蝶 K3 凭证 OCR 填表工具 import os import sys +import json import argparse import yaml import cv2 @@ -31,10 +32,60 @@ PROJECT_DIR = os.path.dirname(os.path.abspath(__file__)) sys.path.insert(0, PROJECT_DIR) from core.recognizer import run_ocr +from core.preprocessor import analyze_image, VoucherPreprocessor from core.parser import K3VoucherParser from core.postprocessor import validate_voucher, validate_merged_vouchers, generate_review_file from core.excel_writer import fill_template, merge_output, merge_voucher_data +# ======== 科目代码标准映射表 ======== +CODE_MAP_PATH = os.path.join(PROJECT_DIR, 'code_map.json') +_code_map_cache = None + +def get_code_map(): + global _code_map_cache + if _code_map_cache is not None: + return _code_map_cache + if os.path.exists(CODE_MAP_PATH): + with open(CODE_MAP_PATH, 'r', encoding='utf-8') as f: + _code_map_cache = json.load(f) + else: + _code_map_cache = {} + return _code_map_cache + +def update_code_map(new_entries: dict): + """更新映射表(用于后续完善)""" + current = get_code_map() + current.update(new_entries) + with open(CODE_MAP_PATH, 'w', encoding='utf-8') as f: + json.dump(current, f, ensure_ascii=False, indent=2) + global _code_map_cache + _code_map_cache = current + +def get_reverse_code_map(): + """名称 → 代码 反向映射(精确匹配)""" + code_map = get_code_map() + return {v: k for k, v in code_map.items()} + +def fuzzy_find_code(name_or_fragment: str): + """根据名称部分字段模糊查找代码 + + 匹配规则:名称包含该字段,或字段包含该名称 + 返回最长的匹配项(最具体) + """ + code_map = get_code_map() + name_or_fragment = name_or_fragment.strip() + if not name_or_fragment or len(name_or_fragment) < 2: + return None + + best_code = None + best_len = 0 + for code, name in code_map.items(): + if name_or_fragment in name or name in name_or_fragment: + if len(name) > best_len: + best_code = code + best_len = len(name) + return best_code + # ======== 路径常量 ======== DEFAULT_CONFIG = os.path.join(PROJECT_DIR, 'config.yaml') @@ -134,6 +185,57 @@ def is_processed(filename): return filename in idx +# ======== PDF 页面提取 ======== +def extract_pdf_pages(pdf_path, output_dir): + """将 PDF 每页转换为 JPG 图片 + + 直接输出到 output_dir,命名格式:_{pdf_name}_p{n}.jpg + 返回 [(page_path, page_label), ...] + """ + import subprocess + import re + + pdf_name = os.path.splitext(os.path.basename(pdf_path))[0] + + # pdftoppm: 每页生成 {prefix}-{页码}.jpg + result = subprocess.run( + ['pdftoppm', '-jpeg', '-r', '300', pdf_path, os.path.join(output_dir, pdf_name)], + capture_output=True, text=True, timeout=120 + ) + if result.returncode != 0: + raise RuntimeError(f"pdftoppm 失败: {result.stderr}") + + # 收集生成的图片 → 重命名为 _{pdf_name}_p{n}.jpg + pages = sorted([ + f for f in os.listdir(output_dir) + if f.startswith(pdf_name) and f.lower().endswith('.jpg') + ]) + + result_list = [] + for p in pages: + old_path = os.path.join(output_dir, p) + # 从文件名提取页码,如 test-01.jpg → _test_p1.jpg + m = re.search(r'-(\d+)\.jpg$', p) + page_num = int(m.group(1)) if m else 0 + new_name = f"_{pdf_name}_p{page_num}.jpg" + new_path = os.path.join(output_dir, new_name) + os.rename(old_path, new_path) + page_label = f"第{page_num}页" if m else os.path.basename(pdf_path) + result_list.append((new_path, page_label)) + + return result_list + + +def cleanup_pdf_temp_files(input_path): + """清理 PDF 提取的临时图片(以下划线开头的 JPG)""" + for f in os.listdir(input_path): + if f.startswith('_') and f.lower().endswith('.jpg'): + try: + os.remove(os.path.join(input_path, f)) + except OSError: + pass + + # ======== 单张凭证处理 ======== def process_single(image_path, config, output_path=None, verbose=False, mark_done=True): """处理单张凭证照片 @@ -153,20 +255,94 @@ def process_single(image_path, config, output_path=None, verbose=False, mark_don if img is None: raise ValueError(f"无法读取图片: {image_path}") - # 2. 原图直出 OCR(无预处理) ocr_config = config.get('ocr', {}) + # 2. 双管线 OCR:原图 + 预处理图,合并结果 if verbose: - print(f" 📋 OCR 识别中...") - lines = run_ocr(img, ocr_config) + print(f" 📊 分析图片质量...") + quality = analyze_image(img) if verbose: - print(f" OCR: {len(lines)} 行文字") + print(f" 📊 照片类型: {quality.photo_type} (亮度={quality.brightness:.0f}, 阴影比={quality.shadow_ratio:.2f})") - # 3. 解析 - parser = K3VoucherParser(lines) + # 管线1: 原图OCR(保留横杠标记/续行结构) + if verbose: + print(f" 📋 管线1: 原图OCR...") + lines1 = run_ocr(img, ocr_config, photo_type=quality.photo_type) + + # 管线2: 预处理后OCR(去阴影/纠偏/锐化,提高文字识别率) + preprocess_config = config.get('preprocessing', {}) + preprocessor = VoucherPreprocessor.from_config(preprocess_config, quality.photo_type) + img_processed = preprocessor.process(img) + + if verbose: + print(f" 📋 管线2: 预处理图OCR...") + lines2 = run_ocr(img_processed, ocr_config, photo_type=quality.photo_type) + + # 合并双管线结果:按 y 坐标邻近去重 + all_lines = lines1 + lines2 + # 按 y 排序后,合并相近位置的重复文本 + all_lines.sort(key=lambda l: (l[3], l[2])) + merged_lines = [] + for line in all_lines: + is_dup = False + for existing in merged_lines: + if abs(line[3] - existing[3]) < 20 and abs(line[2] - existing[2]) < 50: + # 同位置,保留文本更长的 + if len(line[0].strip()) > len(existing[0].strip()): + existing = line + is_dup = True + break + if not is_dup: + merged_lines.append(line) + + if verbose: + print(f" 📋 双管线合并: {len(lines1)}+{len(lines2)} → {len(merged_lines)} 行去重") + + # 3. 解析(使用双管线合并后的数据) + parser = K3VoucherParser(merged_lines) voucher = parser.parse() + # 3.5 双向对照表校正(代码→名称 + 名称→代码反查) + CODE_MAP = get_code_map() + REVERSE_MAP = get_reverse_code_map() + corrections = 0 + for e in voucher['entries']: + code = e.get('account_code', '') + name = e.get('account_name', '') + + # 正向:代码在映射表中 → 用标准名称 + if code in CODE_MAP: + std_name = CODE_MAP[code] + if name != std_name: + if verbose: + print(f" 🔄 名称修正: {code} [{name}] → [{std_name}]") + e['account_name'] = std_name + corrections += 1 + + # 反向:名称精确匹配 → 用标准代码 + if name in REVERSE_MAP: + std_code = REVERSE_MAP[name] + if code != std_code: + if verbose: + print(f" 🔄 代码修正: [{code}] → [{std_code}] (名称={name})") + e['account_code'] = std_code + corrections += 1 + + # 模糊反向:代码不在映射表但名称部分匹配 → 反查 + if code not in CODE_MAP and name: + fuzzy_code = fuzzy_find_code(name) + if fuzzy_code: + std_name = CODE_MAP[fuzzy_code] + if verbose: + print(f" 🔄 模糊反查: [{code},{name}] → [{fuzzy_code},{std_name}]") + e['account_code'] = fuzzy_code + e['account_name'] = std_name + corrections += 1 + + if verbose and corrections > 0: + print(f" ✅ 对照表双向修正: {corrections} 处") + if verbose: print(f" 公司: {voucher['company']}") print(f" 日期: {voucher['date'].strftime('%Y-%m-%d') if voucher['date'] else '未识别'}") @@ -218,6 +394,7 @@ def show_status(config): exts = ('.jpg', '.jpeg', '.png', '.bmp', '.tiff', '.tif', '.webp') pending_files = sorted([f for f in os.listdir(input_dir) if f.lower().endswith(exts)]) if os.path.exists(input_dir) else [] + pdf_files = sorted([f for f in os.listdir(input_dir) if f.lower().endswith('.pdf')]) if os.path.exists(input_dir) else [] done_files = sorted([f for f in os.listdir(done_dir) if f.lower().endswith(exts)]) if os.path.exists(done_dir) else [] processed = set(idx.keys()) @@ -225,7 +402,9 @@ def show_status(config): print("📊 KOCR v4 处理状态") print(f" {'='*40}") print(f" 索引记录: {len(processed)} 张已处理") - print(f" input/ 待处理: {len(pending_files)} 张") + print(f" input/ 待处理: {len(pending_files)} 张图片") + if pdf_files: + print(f" input/ 待转换 PDF: {len(pdf_files)} 个") print(f" done/ 已归档: {len(done_files)} 张") print() @@ -315,17 +494,32 @@ def main(): if os.path.isdir(input_path): # 批量模式 - exts = ('.jpg', '.jpeg', '.png', '.bmp', '.tiff', '.tif', '.webp') - files = sorted([f for f in os.listdir(input_path) if f.lower().endswith(exts)]) - + img_exts = ('.jpg', '.jpeg', '.png', '.bmp', '.tiff', '.tif', '.webp') + files = sorted([f for f in os.listdir(input_path) if f.lower().endswith(img_exts)]) + + # PDF 文件检测与提取 + pdf_files = sorted([f for f in os.listdir(input_path) if f.lower().endswith('.pdf')]) + if pdf_files: + print(f"📄 发现 {len(pdf_files)} 个 PDF 文件,正在提取页面...") + for pdf_file in pdf_files: + pdf_path = os.path.join(input_path, pdf_file) + try: + pages = extract_pdf_pages(pdf_path, input_path) + # 将页面图片加入待处理列表 + page_fnames = [os.path.basename(p) for p, _ in pages] + files.extend(page_fnames) + print(f" 📄 {pdf_file} → {len(pages)} 页") + except Exception as e: + print(f" ❌ {pdf_file} 提取失败: {e}") + if not files: - print(f"📭 {input_path} 中没有图片文件") + print(f"📭 {input_path} 中没有图片或 PDF 文件") sys.exit(0) - - # --new 模式过滤 + + # --new 模式过滤(PDF 临时页面以下划线开头,始终视为新文件) idx = load_index() if args.new or not args.force: - new_files = [f for f in files if f not in idx] + new_files = [f for f in files if f not in idx or f.startswith('_')] skipped = len(files) - len(new_files) files = new_files if skipped > 0: @@ -434,12 +628,15 @@ def main(): for fname, voucher, out_path, quality in results: src = os.path.join(input_path, fname) dst = os.path.join(done_dir, fname) - if os.path.exists(src): + if os.path.exists(src) and not fname.startswith('_'): os.rename(src, dst) archived += 1 if archived > 0: print(f" 📦 已归档 {archived} 张照片到 done/") - + + # ---- 清理 PDF 临时页面图片 ---- + cleanup_pdf_temp_files(input_path) + print(f"🎉 完成! {len(results)}/{len(files)} 成功") else: