feat: 添加调试选项和二次OCR功能，增强文本识别准确性

ocr_tools/universal_doc_parser/models/adapters/mineru_wired_table.py

@@ -3,6 +3,7 @@ import html
 import copy
 from typing import Any, Dict, List, Tuple, Optional, cast
 import ast
+from dataclasses import dataclass
 
 import cv2
 import numpy as np
@@ -28,6 +29,17 @@ class MinerUWiredTableRecognizer:
     - recognize_v4(): 改进流程，使用自定义HTML生成和文本填充（支持data-bbox属性）
     """
 
+    @dataclass
+    class DebugOptions:
+        enabled: bool = False
+        output_dir: Optional[str] = None
+        save_table_lines: bool = False
+        save_connected_components: bool = False
+        save_grid_structure: bool = False
+        save_text_overlay: bool = False
+        image_format: str = "png"
+        prefix: str = ""
+
     def __init__(self, config: Dict[str, Any], ocr_engine: Any):
         self.config = config or {}
         self.upscale_ratio: float = self.config.get("upscale_ratio", 10 / 3)
@@ -41,6 +53,48 @@ class MinerUWiredTableRecognizer:
         self.table_model = UnetTableModel(ocr_engine)
         self.ocr_engine = ocr_engine
 
+        # 统一的调试选项，默认关闭写盘
+        self.debug_options = self._merge_debug_options(self.config.get("debug_options"))
+
+    # ======== Debug 选项与保存工具 ========
+    def _merge_debug_options(self, override: Optional[Dict[str, Any]] = None) -> "MinerUWiredTableRecognizer.DebugOptions":
+        base_dir = None
+        # 兼容旧配置键
+        if isinstance(self.config.get("debug_output_dir"), str):
+            base_dir = self.config.get("debug_output_dir")
+        opts = MinerUWiredTableRecognizer.DebugOptions(
+            enabled=bool(self.config.get("debug_enabled", False)),
+            output_dir=base_dir,
+            save_table_lines=bool(self.config.get("save_table_lines", False)),
+            save_connected_components=bool(self.config.get("save_connected_components", False)),
+            save_grid_structure=bool(self.config.get("save_grid_structure", False)),
+            save_text_overlay=bool(self.config.get("save_text_overlay", False)),
+            image_format=str(self.config.get("debug_image_format", "png")),
+            prefix=str(self.config.get("debug_prefix", "")),
+        )
+        if override and isinstance(override, dict):
+            # 覆盖层允许临时启用或指定目录
+            for k, v in override.items():
+                if hasattr(opts, k):
+                    setattr(opts, k, v)
+        return opts
+
+    def _debug_is_on(self, flag: str, opts: Optional["MinerUWiredTableRecognizer.DebugOptions"] = None) -> bool:
+        o = opts or self.debug_options
+        if not o or not o.enabled:
+            return False
+        if not o.output_dir:
+            return False
+        return bool(getattr(o, flag, False))
+
+    def _debug_path(self, name: str, opts: Optional["MinerUWiredTableRecognizer.DebugOptions"] = None) -> Optional[str]:
+        o = opts or self.debug_options
+        if not o or not o.output_dir:
+            return None
+        prefix = (o.prefix + "_") if o.prefix else ""
+        ext = o.image_format or "png"
+        return f"{o.output_dir}/{prefix}{name}.{ext}"
+
     # ========== 坐标格式转换工具 ==========
     
     @staticmethod
@@ -281,7 +335,7 @@ class MinerUWiredTableRecognizer:
         self,
         bboxes: List[List[float]],
         ocr_boxes: List[Dict[str, Any]],
-    ) -> List[str]:
+    ) -> Tuple[List[str], List[float]]:
         """
         使用中心点落格策略填充文本。
         
@@ -295,11 +349,13 @@ class MinerUWiredTableRecognizer:
             
         Returns:
             每个单元格的文本列表
+            每个单元格的置信度列表
         """
         texts: List[str] = ["" for _ in bboxes]
+        scores: List[float] = [0.0 for _ in bboxes]
         
         if not ocr_boxes:
-            return texts
+            return texts, scores
         
         # 预处理OCR结果：计算中心点
         ocr_items: List[Dict[str, Any]] = []
@@ -314,33 +370,29 @@ class MinerUWiredTableRecognizer:
                 "center_y": cy,
                 "y1": box[1],
                 "text": item.get("text", ""),
-                "confidence": item.get("confidence", 0.0),
+                "confidence": float(item.get("confidence", item.get("score", 1.0))),
             })
         
         # 为每个单元格匹配OCR文本
         for idx, bbox in enumerate(bboxes):
             x1, y1, x2, y2 = bbox
-            matched: List[Tuple[str, float]] = []
+            matched: List[Tuple[str, float, float]] = [] # (text, y1, score)
             
             for ocr in ocr_items:
                 if x1 <= ocr["center_x"] <= x2 and y1 <= ocr["center_y"] <= y2:
-                    matched.append((ocr["text"], ocr["y1"]))
+                    matched.append((ocr["text"], ocr["y1"], ocr["confidence"]))
             
             if matched:
                 # 按y坐标排序，确保多行文本顺序正确
                 matched.sort(key=lambda x: x[1])
-                texts[idx] = " ".join([t for t, _ in matched])
+                texts[idx] = "".join([t for t, _, _ in matched])
+                # 计算平均置信度
+                avg_score = sum([s for _, _, s in matched]) / len(matched)
+                scores[idx] = avg_score
+            else:
+                scores[idx] = 0.0 # 无匹配文本，置信度为0
         
-        return texts
-    
-    def _match_text_by_center(
-        self,
-        cells_bbox: List[List[float]],
-        ocr_boxes: List[Dict[str, Any]],
-    ) -> List[str]:
-        """使用中心点落格分配文本，行内按 y 排序后拼接。（旧版兼容）"""
-        return self._fill_text_by_center_point(cells_bbox, ocr_boxes)
-    
+        return texts, scores
     
     def recognize_legacy(
         self,
@@ -401,41 +453,12 @@ class MinerUWiredTableRecognizer:
             # 缩回裁剪坐标
             polys /= upscale
             bboxes = [self._poly_to_bbox(poly) for poly in polys]
-            texts = self._match_text_by_center(bboxes, ocr_boxes or [])
-            # 对空文本单元格触发单元格级 OCR 补充
+            texts, scores = self._fill_text_by_center_point(bboxes, ocr_boxes)
+
+            # 统一调用二次OCR封装（替换原有重复逻辑）
             if self.ocr_engine is not None and any(not t for t in texts):
-                crop_list = []
-                crop_info = []
-                h, w = table_image.shape[:2]
-                margin = self.cell_crop_margin
-                for idx, bbox in enumerate(bboxes):
-                    if texts[idx]:
-                        continue
-                    x1, y1, x2, y2 = bbox
-                    x1i, y1i, x2i, y2i = map(int, [x1, y1, x2, y2])
-                    
-                    # 增加裁剪边距防止文字被截断（特别是边界字符如"司"）
-                    x1i = max(0, x1i - margin)
-                    y1i = max(0, y1i - margin)
-                    x2i = min(w, x2i + margin)
-                    y2i = min(h, y2i + margin)
-                    
-                    if x2i <= x1i or y2i <= y1i:
-                        continue
-                    crop = table_image[y1i:y2i, x1i:x2i]
-                    if crop.size == 0:
-                        continue
-                    crop_list.append(crop)
-                    crop_info.append(idx)
-                if crop_list:
-                    try:
-                        ocr_res = self.ocr_engine.ocr(crop_list, det=False)
-                        if ocr_res and isinstance(ocr_res, list) and len(ocr_res) == 1:
-                            for loc, (text, score) in zip(crop_info, ocr_res[0]):
-                                if score >= self.ocr_conf_threshold and text:
-                                    texts[loc] = text
-                    except Exception:
-                        pass
+                texts = self._second_pass_ocr_fill(table_image, bboxes, texts, scores)
+
             for idx, bbox in enumerate(bboxes):
                 lp = logic_points[idx] if len(logic_points) > idx else [0, 0, 0, 0]
                 cells.append({
@@ -449,12 +472,11 @@ class MinerUWiredTableRecognizer:
 
         # 通过BeautifulSoup增强HTML，添加data-bbox和data-score属性（保留原始HTML结构）
         html_enhanced = self._enhance_html_with_cell_data(html_code, cells)
-
         return {
             "html": html_enhanced or html_code or "",
             "cells": cells,
         }
-    
+
     def _enhance_html_with_cell_data(self, html_code: str, cells: List[Dict[str, Any]]) -> str:
         """通过BeautifulSoup增强HTML，为每个td添加data-bbox和data-score属性
         
@@ -527,7 +549,6 @@ class MinerUWiredTableRecognizer:
         hpred_up: np.ndarray,
         vpred_up: np.ndarray,
         upscale: float = 1.0,
-        debug_output_dir: Optional[str] = None
     ) -> List[List[float]]:
         """
         基于连通域分析从表格线 Mask 提取单元格
@@ -573,6 +594,15 @@ class MinerUWiredTableRecognizer:
             # 过滤掉长条形的非单元格区域（例如边缘的细长空白）
             if w_cell > w * 0.95 or h_cell > h * 0.95:
                 continue
+
+            # 转换到原图尺度
+            orig_h = h_cell / upscale
+            orig_w = w_cell / upscale
+            
+            # 过滤极小高度/宽度的单元格 (可能是边缘噪声或线条残留)
+            # 阈值设为 5 像素，通常文本行不会小于这个高度
+            if orig_h < 5 or orig_w < 5:
+                continue
                 
             # 还原到原图坐标
             # 注意：连通域提取的是内部空白，实际单元格边界应该包含线条的一半宽度
@@ -589,18 +619,6 @@ class MinerUWiredTableRecognizer:
         bboxes.sort(key=lambda b: (int(b[1] / 10), b[0]))
         
         logger.info(f"连通域分析提取到 {len(bboxes)} 个单元格")
-        
-        # 调试可视化
-        if debug_output_dir:
-            vis = np.zeros((h, w, 3), dtype=np.uint8)
-            vis[grid_mask > 0] = [0, 0, 255] # 红色线条
-            
-            # 绘制提取出的框
-            for i, box in enumerate(bboxes):
-                x1, y1, x2, y2 = [int(c * upscale) for c in box]
-                cv2.rectangle(vis, (x1, y1), (x2, y2), (0, 255, 0), 2)
-                
-            cv2.imwrite(f"{debug_output_dir}/connected_components.png", vis)
             
         return bboxes
     
@@ -643,7 +661,111 @@ class MinerUWiredTableRecognizer:
         cv2.imwrite(output_path, vis_img)
         logger.info(f"检测线可视化: {output_path}")
     
+    def _visualize_connected_components(
+        self,
+        hpred_up: np.ndarray,
+        vpred_up: np.ndarray,
+        bboxes: List[List[float]],
+        upscale: float,
+        output_path: str
+    ) -> None:
+        """
+        复刻连通域风格：红色网格线背景 + 绿色单元格框。
+        使用上采样尺度的 mask 与坐标，保证线条清晰。
+        """
+        h, w = hpred_up.shape[:2]
+
+        # 与连通域提取相同的预处理，以获得直观的网格线背景
+        _, h_bin = cv2.threshold(hpred_up, 127, 255, cv2.THRESH_BINARY)
+        _, v_bin = cv2.threshold(vpred_up, 127, 255, cv2.THRESH_BINARY)
+        kernel_h = cv2.getStructuringElement(cv2.MORPH_RECT, (5, 1))
+        kernel_v = cv2.getStructuringElement(cv2.MORPH_RECT, (1, 5))
+        h_bin = cv2.dilate(h_bin, kernel_h, iterations=1)
+        v_bin = cv2.dilate(v_bin, kernel_v, iterations=1)
+        grid_mask = cv2.bitwise_or(h_bin, v_bin)
+
+        vis = np.zeros((h, w, 3), dtype=np.uint8)
+        vis[grid_mask > 0] = [0, 0, 255]  # 红色线条
+
+        # 在上采样坐标系上绘制单元格框
+        for box in bboxes:
+            x1, y1, x2, y2 = [int(c * upscale) for c in box]
+            cv2.rectangle(vis, (x1, y1), (x2, y2), (0, 255, 0), 2)
+
+        cv2.imwrite(output_path, vis)
+        logger.info(f"连通域可视化: {output_path}")
     
+    def _compress_grid(self, cells: List[Dict]) -> List[Dict]:
+        """
+        压缩网格索引，移除空行和空列
+        """
+        if not cells:
+            return []
+            
+        # 1. 计算当前最大行列
+        max_row = 0
+        max_col = 0
+        for cell in cells:
+            max_row = max(max_row, cell["row"] + cell.get("rowspan", 1))
+            max_col = max(max_col, cell["col"] + cell.get("colspan", 1))
+            
+        # 2. 标记占用情况
+        # 修改：仅标记有单元格起始的行/列为占用
+        # 这样可以自动移除那些仅由 rowspan/colspan 跨越的"间隙行/列"
+        # 例如：表头行之间的微小垂直间隙，如果没有单元格从该间隙开始，则该行应被折叠
+        row_occupied = [False] * max_row
+        col_occupied = [False] * max_col
+        
+        for cell in cells:
+            if cell["row"] < max_row:
+                row_occupied[cell["row"]] = True
+            if cell["col"] < max_col:
+                col_occupied[cell["col"]] = True
+                    
+        # 3. 构建映射表
+        row_map = [0] * (max_row + 1)
+        current_row = 0
+        for r in range(max_row):
+            if row_occupied[r]:
+                current_row += 1
+            row_map[r + 1] = current_row
+            
+        col_map = [0] * (max_col + 1)
+        current_col = 0
+        for c in range(max_col):
+            if col_occupied[c]:
+                current_col += 1
+            col_map[c + 1] = current_col
+            
+        # 4. 更新单元格索引
+        new_cells = []
+        for cell in cells:
+            new_cell = cell.copy()
+            
+            old_r1 = cell["row"]
+            old_r2 = old_r1 + cell.get("rowspan", 1)
+            new_r1 = row_map[old_r1]
+            new_r2 = row_map[old_r2]
+            
+            old_c1 = cell["col"]
+            old_c2 = old_c1 + cell.get("colspan", 1)
+            new_c1 = col_map[old_c1]
+            new_c2 = col_map[old_c2]
+            
+            # 如果压缩后 span 变为 0 (理论上不应该，因为只要有 cell 占用，occupied 就是 True)，
+            # 但为了安全起见，确保至少为 1
+            new_span_r = max(1, new_r2 - new_r1)
+            new_span_c = max(1, new_c2 - new_c1)
+            
+            new_cell["row"] = new_r1
+            new_cell["col"] = new_c1
+            new_cell["rowspan"] = new_span_r
+            new_cell["colspan"] = new_span_c
+            
+            new_cells.append(new_cell)
+            
+        return new_cells
+
     def _recover_grid_structure(self, bboxes: List[List[float]]) -> List[Dict]:
         """
         从散乱的单元格 bbox 恢复表格的行列结构 (row, col, rowspan, colspan)
@@ -726,6 +848,9 @@ class MinerUWiredTableRecognizer:
         # 按行列排序
         structured_cells.sort(key=lambda c: (c["row"], c["col"]))
         
+        # 压缩网格，移除空行空列
+        structured_cells = self._compress_grid(structured_cells)
+        
         return structured_cells
 
     def _build_html_from_merged_cells(self, merged_cells: List[Dict]) -> str:
@@ -778,7 +903,7 @@ class MinerUWiredTableRecognizer:
                     if rowspan > 1:
                         attrs.append(f'rowspan="{rowspan}"')
                     
-                    html_parts.append(f'<td {" ".join(attrs)}>{text}</td>')
+                    html_parts.append(f'<td {"".join(attrs)}>{text}</td>')
                     
                     # 标记占用
                     for i in range(rowspan):
@@ -836,7 +961,7 @@ class MinerUWiredTableRecognizer:
         self,
         table_image: np.ndarray,
         ocr_boxes: List[Dict[str, Any]],
-        debug_output_dir: Optional[str] = None,
+        debug_options: Optional[Dict[str, Any]] = None,
     ) -> Dict[str, Any]:
         """
         V4版本：直接从表格线计算单元格，绕过 MinerU 的 cal_region_boxes
@@ -862,117 +987,60 @@ class MinerUWiredTableRecognizer:
         hpred_up = cv2.resize(hpred, (w_up, h_up), interpolation=cv2.INTER_NEAREST)
         vpred_up = cv2.resize(vpred, (w_up, h_up), interpolation=cv2.INTER_NEAREST)
         
+        # 调试选项合并
+        dbg = self._merge_debug_options(debug_options or {})
+
         # Step 1.5: 可视化表格线（调试用）- 需要缩放回原图
-        if debug_output_dir:
+        if self._debug_is_on("save_table_lines", dbg):
             hpred_orig = cv2.resize(hpred_up, (w, h), interpolation=cv2.INTER_NEAREST)
             vpred_orig = cv2.resize(vpred_up, (w, h), interpolation=cv2.INTER_NEAREST)
+            out_path = self._debug_path("unet_table_lines", dbg)
             self._visualize_table_lines(
                 table_image,
                 hpred_orig,
                 vpred_orig,
-                output_path=f"{debug_output_dir}/unet_table_lines.png"
+                output_path=out_path if out_path else ""
             )
         
         # Step 2: 使用连通域法提取单元格 (替换了原来的投影法)
-        bboxes = self._compute_cells_from_lines(hpred_up, vpred_up, upscale, debug_output_dir)
-        
+        bboxes = self._compute_cells_from_lines(hpred_up, vpred_up, upscale)
         if not bboxes:
             raise RuntimeError("未能提取出单元格")
 
+        # Step 2.5: 可视化连通域（线条+框，直观版）
+        if self._debug_is_on("save_connected_components", dbg):
+            out_path = self._debug_path("connected_components", dbg)
+            if out_path:
+                self._visualize_connected_components(
+                    hpred_up,
+                    vpred_up,
+                    bboxes,
+                    upscale,
+                    output_path=out_path
+                )
+
         # Step 3: 重建网格结构 (计算 row, col, rowspan, colspan)
         # 这一步替代了原来的 _merge_cells_without_separator
         merged_cells = self._recover_grid_structure(bboxes)
         
         # Step 3.5: 可视化逻辑结构 (新增)
-        if debug_output_dir:
-            self._visualize_grid_structure(
-                table_image, merged_cells,
-                output_path=f"{debug_output_dir}/grid_structure.png"
-            )
+        if self._debug_is_on("save_grid_structure", dbg):
+            out_path = self._debug_path("grid_structure", dbg)
+            if out_path:
+                self._visualize_grid_structure(
+                    table_image, merged_cells,
+                    output_path=out_path
+                )
         
         # Step 4: 统一计算文本填充
         bboxes_merged = [cell["bbox"] for cell in merged_cells]
-        texts = self._fill_text_by_center_point(bboxes_merged, ocr_boxes or [])
+        texts, scores = self._fill_text_by_center_point(bboxes_merged, ocr_boxes or [])
         
-        # Step 4.5: 对空单元格尝试二次 OCR (新增)
-        # 针对漏检问题（特别是竖排小字），进行切片放大识别
-        if hasattr(self, 'ocr_engine') and self.ocr_engine and any(not t for t in texts):
-            crop_list = []
-            crop_indices = []
-            h_img, w_img = table_image.shape[:2]
-            margin = self.cell_crop_margin
-            
-            for i, text in enumerate(texts):
-                if text.strip():
-                    continue
-                
-                bbox = bboxes_merged[i]
-                x1, y1, x2, y2 = map(int, bbox)
-                
-                # 边界保护 + 少量外扩
-                x1 = max(0, x1 - margin)
-                y1 = max(0, y1 - margin)
-                x2 = min(w_img, x2 + margin)
-                y2 = min(h_img, y2 + margin)
-                
-                if x2 <= x1 or y2 <= y1:
-                    continue
-                    
-                cell_img = table_image[y1:y2, x1:x2]
-                if cell_img.size == 0:
-                    continue
-                
-                # --- 关键改进：放大与旋转 ---
-                cell_h, cell_w = cell_img.shape[:2]
-                
-                # 1. 放大图像：对于表格中的小字，放大能显著提高识别率
-                # 建议放大 2 倍，如果原图特别小可以更大
-                scale = 2.0
-                if cell_h < 64 or cell_w < 64: # 只有较小的图才放大，避免大图过大
-                     cell_img = cv2.resize(cell_img, None, fx=scale, fy=scale, interpolation=cv2.INTER_CUBIC)
-                
-                # 2. 处理竖排文本：如果高宽比很大（>2），很可能是竖排表头（如"优先股"）
-                # 通用 OCR 模型通常只支持横排，旋转 90 度变成横排
-                if cell_h > cell_w * 2:
-                    cell_img = cv2.rotate(cell_img, cv2.ROTATE_90_COUNTERCLOCKWISE)
-                # -------------------------
-                
-                crop_list.append(cell_img)
-                crop_indices.append(i)
-            
-            if crop_list:
-                try:
-                    # 批量识别，det=False 表示直接识别内容（假设裁剪图就是文本行）
-                    ocr_res = self.ocr_engine.ocr(crop_list, det=False)
-                    
-                    # 解析结果 (兼容 PaddleOCR 返回格式)
-                    # ocr_res 结构通常为 [(text, score), (text, score), ...] 对应每张图
-                    # 但有时可能包裹在列表中，需做兼容处理
-                    results = ocr_res
-                    if isinstance(ocr_res, list) and len(ocr_res) == 1 and isinstance(ocr_res[0], list) and len(ocr_res[0]) == len(crop_list):
-                         # 兼容 legacy 代码中遇到的 [[(t,s), (t,s)...]] 情况
-                         results = ocr_res[0]
-                    
-                    if len(results) == len(crop_list):
-                        for idx, res in enumerate(results):
-                            # res 可能是 (text, score) 或 [(text, score)] 或 None
-                            if not res: continue
-                            
-                            text = ""
-                            score = 0.0
-                            
-                            if isinstance(res, tuple):
-                                text, score = res
-                            elif isinstance(res, list) and len(res) > 0:
-                                text, score = res[0]
-                            
-                            if score >= self.ocr_conf_threshold and text:
-                                texts[crop_indices[idx]] = text
-                                
-                except Exception as e:
-                    logger.warning(f"二次OCR失败: {e}")
+        # Step 4.5: 二次 OCR 修正
+        # 针对漏检（空文本）、低置信度、竖排文本进行二次识别
+        if hasattr(self, 'ocr_engine') and self.ocr_engine:
+            texts = self._second_pass_ocr_fill(table_image, bboxes_merged, texts, scores)
 
-        # 将文本填入 merged_cells
         for i, cell in enumerate(merged_cells):
             cell["text"] = texts[i] if i < len(texts) else ""
         
@@ -980,11 +1048,13 @@ class MinerUWiredTableRecognizer:
         html_filled = self._build_html_from_merged_cells(merged_cells)
         
         # Step 6: 可视化文本填充（调试用）
-        if debug_output_dir:
-            self._visualize_with_text(
-                table_image, bboxes_merged, texts,
-                output_path=f"{debug_output_dir}/text_filled_v4.png"
-            )
+        if self._debug_is_on("save_text_overlay", dbg):
+            out_path = self._debug_path("text_filled_v4", dbg)
+            if out_path:
+                self._visualize_with_text(
+                    table_image, bboxes_merged, texts,
+                    output_path=out_path
+                )
         
         # Step 7: 组装 cells 输出
         cells = []
@@ -999,7 +1069,7 @@ class MinerUWiredTableRecognizer:
                 "matched_text": cell["text"],
                 "score": 100.0,
             })
-        
+
         return {
             "html": html_filled,
             "cells": cells,
@@ -1148,9 +1218,147 @@ class MinerUWiredTableRecognizer:
         """
         if self.use_custom_postprocess:
             try:
-                return self.recognize_v4(table_image, ocr_boxes, debug_output_dir="./output")
+                return self.recognize_v4(table_image, ocr_boxes, debug_options=self.debug_options.__dict__)
             except Exception:
                 # 回退
                 return self.recognize_legacy(table_image, ocr_boxes)
         else:
-            return self.recognize_legacy(table_image, ocr_boxes)
+            return self.recognize_legacy(table_image, ocr_boxes)
+    
+    def _second_pass_ocr_fill(
+        self,
+        table_image: np.ndarray,
+        bboxes: List[List[float]],
+        texts: List[str],
+        scores: Optional[List[float]] = None,
+    ) -> List[str]:
+        """
+        二次OCR统一封装：
+        - 对空文本单元格裁剪图块并少量外扩
+        - 对低置信度文本进行重识别
+        - 对竖排单元格（高宽比大）进行旋转后识别
+        """
+        try:
+            if not hasattr(self, "ocr_engine") or self.ocr_engine is None:
+                return texts
+            
+            # 如果没有传入 scores，则默认全为 1.0（仅处理空文本）
+            if scores is None:
+                scores = [1.0 if t else 0.0 for t in texts]
+
+            h_img, w_img = table_image.shape[:2]
+            margin = getattr(self, "cell_crop_margin", 2)
+            
+            # 触发二次OCR的阈值
+            trigger_score_thresh = 0.90 
+
+            crop_list: List[np.ndarray] = []
+            crop_indices: List[int] = []
+
+            # 收集需要二次OCR的裁剪块
+            for i, t in enumerate(texts):
+                bbox = bboxes[i]
+                w_box = bbox[2] - bbox[0]
+                h_box = bbox[3] - bbox[1]
+                
+                # 判断是否需要二次OCR
+                need_reocr = False
+                
+                # 1. 文本为空
+                if not t or not t.strip():
+                    need_reocr = True
+                # 2. 置信度过低
+                elif scores[i] < trigger_score_thresh:
+                    need_reocr = True
+                # 3. 竖排单元格 (高宽比 > 2.5) 且置信度不是极高
+                # 竖排文本全图OCR容易出错，旋转后识别更准
+                elif h_box > w_box * 2.5 and scores[i] < 0.98:
+                    need_reocr = True
+
+                if not need_reocr:
+                    continue
+                
+                if i >= len(bboxes):
+                    continue
+
+                x1, y1, x2, y2 = map(int, bboxes[i])
+                x1 = max(0, x1 - margin)
+                y1 = max(0, y1 - margin)
+                x2 = min(w_img, x2 + margin)
+                y2 = min(h_img, y2 + margin)
+                if x2 <= x1 or y2 <= y1:
+                    continue
+
+                cell_img = table_image[y1:y2, x1:x2]
+                if cell_img.size == 0:
+                    continue
+
+                ch, cw = cell_img.shape[:2]
+                # 小图放大
+                if ch < 64 or cw < 64:
+                    cell_img = cv2.resize(cell_img, None, fx=2.0, fy=2.0, interpolation=cv2.INTER_CUBIC)
+                    ch, cw = cell_img.shape[:2]
+
+                # 竖排文本旋转为横排
+                # 这里的阈值设为 2.0，涵盖大部分竖排表头
+                if ch > cw * 2.0:
+                    cell_img = cv2.rotate(cell_img, cv2.ROTATE_90_COUNTERCLOCKWISE)
+
+                crop_list.append(cell_img)
+                crop_indices.append(i)
+
+            if not crop_list:
+                return texts
+            
+            logger.info(f"触发二次OCR: {len(crop_list)} 个单元格 (总数 {len(texts)})")
+
+            # 批量OCR（det=False）
+            ocr_res = self.ocr_engine.ocr(crop_list, det=False)
+            results = ocr_res
+
+            # 兼容 [[(text,score),...]] 的嵌套返回
+            if isinstance(results, list) and len(results) == 1 and isinstance(results[0], list) and len(results[0]) == len(crop_list):
+                results = results[0]
+
+            # 解析为 (text, score)
+            def _parse_item(res_item) -> Tuple[str, float]:
+                if res_item is None:
+                    return "", 0.0
+                # 直接 (text, score)
+                if isinstance(res_item, tuple) and len(res_item) >= 2:
+                    return str(res_item[0] or ""), float(res_item[1] or 0.0)
+                # 列表形式，取第一个
+                if isinstance(res_item, list) and len(res_item) > 0:
+                    first = res_item[0]
+                    if isinstance(first, tuple) and len(first) >= 2:
+                        return str(first[0] or ""), float(first[1] or 0.0)
+                    if isinstance(first, list) and len(first) >= 2:
+                        return str(first[0] or ""), float(first[1] or 0.0)
+                    if isinstance(first, dict):
+                        txt = str(first.get("text") or first.get("label") or "")
+                        sc = float(first.get("score") or first.get("confidence") or 0.0)
+                        return txt, sc
+                # 字典形式
+                if isinstance(res_item, dict):
+                    txt = str(res_item.get("text") or res_item.get("label") or "")
+                    sc = float(res_item.get("score") or res_item.get("confidence") or 0.0)
+                    return txt, sc
+                return "", 0.0
+
+            # 对齐长度，避免越界
+            n = min(len(results) if isinstance(results, list) else 0, len(crop_list), len(crop_indices))
+            # 结果采纳阈值：二次识别的结果如果置信度太低，可能不如不填（或者保留原值？）
+            # 这里策略是：只要有结果且置信度尚可，就覆盖。
+            # 注意：如果原文本不为空但置信度低，二次识别结果置信度更低，是否覆盖？
+            # 目前逻辑是只要 > conf_th 就覆盖。
+            conf_th = float(getattr(self, "ocr_conf_threshold", 0.5))
+
+            for k in range(n):
+                text_k, score_k = _parse_item(results[k])
+                if text_k and score_k >= conf_th:
+                    texts[crop_indices[k]] = text_k
+
+        except Exception as e:
+            logger.warning(f"二次OCR失败: {e}")
+
+        return texts