test

PLAN.md

@@ -0,0 +1,186 @@
+# PDF 處理雙軌制改善計劃 (修訂版 v5)
+
+## 問題分析
+
+### 一、Direct Track 表格問題
+
+| 指標 | edit.pdf | edit3.pdf |
+|------|----------|-----------|
+| 原始表格結構 | 6 rows x 2 cols | 12 rows x 17 cols |
+| PyMuPDF 識別的 cells | 12 (無合併) | **83** (有121個合併) |
+| Direct Track 提取的 cells | 12 | **204** (全部視為1x1) |
+| 跨欄/跨行識別 | 不需要 | **❌ 完全未識別** |
+| 渲染結果 | ✓ 完美 | ❌ 欄位切分錯誤、文字超出 |
+
+**根因**: `_detect_tables_by_position()` 無法識別合併單元格
+
+### 二、Direct Track 圖片問題 (edit3.pdf)
+
+| 問題 | 數量 | 說明 |
+|------|------|------|
+| 極小裝飾圖片 | 3 | < 200 px²，應過濾 |
+| 覆蓋圖像 (黑框) | 6 | 已檢測但未從渲染中移除 |
+| 大型 vector_graphics | 3 | ✓ 已正確過濾 |
+
+### 三、OCR Track 表格問題
+
+| 表格 | cells | cell_boxes | cell_boxes 坐標檢查 |
+|------|-------|------------|-------------------|
+| pp3_0_3 | 13 | 13 | ⚠️ 1/5 超出範圍 |
+| pp3_0_6 | 29 | 12 | ❌ 全部超出範圍 |
+| pp3_0_7 | 12 | 51 | ❌ 全部超出範圍 |
+| pp3_0_16 | 51 | 29 | ❌ 全部超出範圍 |
+
+**根因**: PP-StructureV3 的 cell_boxes 座標系統錯亂
+
+### 四、OCR Track 圖片問題 ❌ 嚴重
+
+| 文件 | 圖片元素 | PP-Structure 原始數據 | 轉換後 UnifiedDocument | 結果 |
+|------|---------|---------------------|----------------------|------|
+| edit.pdf | pp3_1_8 | saved_path="pp3_1_8.png" ✓ | content=字符串 ❌ | 圖片未放回 |
+| edit3.pdf | pp3_1_2 | saved_path="pp3_1_2.png" ✓ | content=字符串 ❌ | 圖片未放回 |
+
+**根因**: `ocr_to_unified_converter.py` 的 `_convert_pp3_element` 方法中：
+
+```python
+# 當前代碼 (第604-613行)
+elif element_type in [ElementType.IMAGE, ElementType.FIGURE]:
+    content = {'path': elem_data.get('img_path', ''), ...}
+else:
+    content = elem_data.get('content', '')  # ← CHART 類型走這裡！
+```
+
+**問題**:
+1. `CHART` 類型未被視為視覺元素
+2. `saved_path` 完全丟失
+3. `content` 變成文字而非圖片路徑
+
+---
+
+## 改善計劃
+
+### 階段 1: Direct Track 使用 PyMuPDF find_tables (優先級：最高)
+
+**問題**: `_detect_tables_by_position` 無法識別合併單元格
+
+**方案**: 改用 PyMuPDF 的 `find_tables()` API
+
+**檔案**: `backend/app/services/direct_extraction_engine.py`
+
+```python
+def _extract_tables_with_pymupdf(self, page, page_num, counter):
+    tables = page.find_tables()
+    for table in tables.tables:
+        # 獲取 cells，保留合併信息
+        cells = []
+        for row_idx in range(table.row_count):
+            for col_idx in range(table.col_count):
+                cell_data = table.cells[row_idx * table.col_count + col_idx]
+                if cell_data is None:
+                    continue  # 跳過被合併的單元格
+                # 計算 row_span/col_span...
+```
+
+### 階段 2: 修復 OCR Track 圖片路徑丟失 (優先級：最高)
+
+**問題**: CHART 類型的 saved_path 在轉換時丟失
+
+**檔案**: `backend/app/services/ocr_to_unified_converter.py`
+**位置**: `_convert_pp3_element` 方法，約第604行
+
+**修改**:
+
+```python
+# 修改前
+elif element_type in [ElementType.IMAGE, ElementType.FIGURE]:
+
+# 修改後：包含所有視覺元素類型
+elif element_type in [
+    ElementType.IMAGE, ElementType.FIGURE, ElementType.CHART,
+    ElementType.DIAGRAM, ElementType.LOGO, ElementType.STAMP
+]:
+    # 優先使用 saved_path
+    image_path = (
+        elem_data.get('saved_path') or
+        elem_data.get('img_path') or
+        ''
+    )
+    content = {
+        'saved_path': image_path,  # 關鍵：保留 saved_path
+        'path': image_path,
+        'width': elem_data.get('width', 0),
+        'height': elem_data.get('height', 0),
+        'format': elem_data.get('format', 'unknown')
+    }
+```
+
+### 階段 3: 修復 OCR Track cell_boxes 座標 (優先級：高)
+
+**方案**: 驗證座標，超出範圍時使用 CV 線檢測 fallback
+
+### 階段 4: 過濾極小裝飾圖片 (優先級：高)
+
+```python
+if elem_area < 200:
+    continue  # 跳過 < 200 px² 的圖片
+```
+
+### 階段 5: 過濾覆蓋圖像 (優先級：高)
+
+在提取階段過濾與 covering_images 重疊的圖片。
+
+---
+
+## 實施優先級
+
+| 階段 | 描述 | 優先級 | 影響 |
+|------|------|--------|------|
+| 1 | Direct Track 使用 PyMuPDF find_tables | **最高** | 修復合併單元格 |
+| 2 | **OCR Track 圖片路徑修復** | **最高** | 修復圖片未放回 |
+| 3 | OCR Track cell_boxes 座標修復 | 高 | 修復表格渲染錯亂 |
+| 4 | 過濾極小裝飾圖片 | 高 | 減少無意義圖片 |
+| 5 | 過濾覆蓋圖像 | 高 | 減少黑框 |
+
+---
+
+## 預期效果
+
+### Direct Track
+
+| 指標 | 修改前 | 修改後 |
+|------|--------|--------|
+| edit3.pdf cells | 204 (錯誤拆分) | 83 (正確識別合併) |
+| 跨欄/跨行識別 | ❌ | ✓ |
+
+### OCR Track 圖片
+
+| 指標 | 修改前 | 修改後 |
+|------|--------|--------|
+| pp3_1_8 (edit.pdf) | 圖片未放回 | ✓ 正確放回 |
+| pp3_1_2 (edit3.pdf) | 圖片未放回 | ✓ 正確放回 |
+
+### OCR Track 表格
+
+| 指標 | 修改前 | 修改後 |
+|------|--------|--------|
+| cell_boxes 座標 | 3/5 表格錯誤 | 全部正確或 CV fallback |
+
+---
+
+## 測試計劃
+
+1. **edit.pdf Direct Track**: 確保無回歸
+
+2. **edit3.pdf Direct Track**:
+   - 驗證表格識別到 83 cells（非 204）
+   - 驗證跨欄/跨行正確
+   - 驗證極小圖片被過濾
+   - 驗證黑框被過濾
+
+3. **edit.pdf OCR Track**:
+   - **驗證 pp3_1_8.png 正確放回**
+   - 驗證 cell_boxes 座標修復
+
+4. **edit3.pdf OCR Track**:
+   - **驗證 pp3_1_2.png 正確放回**
+   - 驗證 cell_boxes 座標修復

backend/app/services/ocr_to_unified_converter.py

@@ -178,6 +178,114 @@ def trim_empty_columns(table_dict: Dict[str, Any]) -> Dict[str, Any]:
     return result
 
 
+def validate_cell_boxes(
+    cell_boxes: List[List[float]],
+    table_bbox: List[float],
+    page_width: float,
+    page_height: float,
+    tolerance: float = 5.0
+) -> Dict[str, Any]:
+    """
+    Validate cell_boxes coordinates against page boundaries and table bbox.
+
+    PP-StructureV3 sometimes returns cell_boxes with coordinates that exceed
+    page boundaries. This function validates and reports issues.
+
+    Args:
+        cell_boxes: List of cell bounding boxes [[x0, y0, x1, y1], ...]
+        table_bbox: Table bounding box [x0, y0, x1, y1]
+        page_width: Page width in pixels
+        page_height: Page height in pixels
+        tolerance: Allowed tolerance for boundary checks (pixels)
+
+    Returns:
+        Dict with:
+            - valid: bool - whether all cell_boxes are valid
+            - invalid_count: int - number of invalid cell_boxes
+            - clamped_boxes: List - cell_boxes clamped to valid boundaries
+            - issues: List[str] - description of issues found
+    """
+    if not cell_boxes:
+        return {'valid': True, 'invalid_count': 0, 'clamped_boxes': [], 'issues': []}
+
+    issues = []
+    invalid_count = 0
+    clamped_boxes = []
+
+    # Page boundaries with tolerance
+    min_x = -tolerance
+    min_y = -tolerance
+    max_x = page_width + tolerance
+    max_y = page_height + tolerance
+
+    for idx, box in enumerate(cell_boxes):
+        if not box or len(box) < 4:
+            issues.append(f"Cell {idx}: Invalid box format")
+            invalid_count += 1
+            clamped_boxes.append([0, 0, 0, 0])
+            continue
+
+        x0, y0, x1, y1 = box[:4]
+        is_valid = True
+        cell_issues = []
+
+        # Check if coordinates exceed page boundaries
+        if x0 < min_x:
+            cell_issues.append(f"x0={x0:.1f} < 0")
+            is_valid = False
+        if y0 < min_y:
+            cell_issues.append(f"y0={y0:.1f} < 0")
+            is_valid = False
+        if x1 > max_x:
+            cell_issues.append(f"x1={x1:.1f} > page_width={page_width:.1f}")
+            is_valid = False
+        if y1 > max_y:
+            cell_issues.append(f"y1={y1:.1f} > page_height={page_height:.1f}")
+            is_valid = False
+
+        # Check for inverted coordinates
+        if x0 > x1:
+            cell_issues.append(f"x0={x0:.1f} > x1={x1:.1f}")
+            is_valid = False
+        if y0 > y1:
+            cell_issues.append(f"y0={y0:.1f} > y1={y1:.1f}")
+            is_valid = False
+
+        if not is_valid:
+            invalid_count += 1
+            issues.append(f"Cell {idx}: {', '.join(cell_issues)}")
+
+        # Clamp to valid boundaries
+        clamped_box = [
+            max(0, min(x0, page_width)),
+            max(0, min(y0, page_height)),
+            max(0, min(x1, page_width)),
+            max(0, min(y1, page_height))
+        ]
+
+        # Ensure proper ordering after clamping
+        if clamped_box[0] > clamped_box[2]:
+            clamped_box[0], clamped_box[2] = clamped_box[2], clamped_box[0]
+        if clamped_box[1] > clamped_box[3]:
+            clamped_box[1], clamped_box[3] = clamped_box[3], clamped_box[1]
+
+        clamped_boxes.append(clamped_box)
+
+    if invalid_count > 0:
+        logger.warning(
+            f"Cell boxes validation: {invalid_count}/{len(cell_boxes)} invalid. "
+            f"Page: {page_width:.0f}x{page_height:.0f}, Table bbox: {table_bbox}"
+        )
+
+    return {
+        'valid': invalid_count == 0,
+        'invalid_count': invalid_count,
+        'clamped_boxes': clamped_boxes,
+        'issues': issues,
+        'needs_fallback': invalid_count > len(cell_boxes) * 0.5  # >50% invalid = needs fallback
+    }
+
+
 class OCRToUnifiedConverter:
     """
     Converter for transforming PP-StructureV3 OCR results to UnifiedDocument format.
@@ -337,19 +445,22 @@ class OCRToUnifiedConverter:
         for page_idx, page_result in enumerate(enhanced_results):
             elements = []
 
+            # Get page dimensions first (needed for element conversion)
+            page_width = page_result.get('width', 0)
+            page_height = page_result.get('height', 0)
+            pp_dimensions = Dimensions(width=page_width, height=page_height)
+
             # Process elements from parsing_res_list
             if 'elements' in page_result:
                 for elem_data in page_result['elements']:
-                    element = self._convert_pp3_element(elem_data, page_idx)
+                    element = self._convert_pp3_element(
+                        elem_data, page_idx,
+                        page_width=page_width,
+                        page_height=page_height
+                    )
                     if element:
                         elements.append(element)
 
-            # Get page dimensions
-            pp_dimensions = Dimensions(
-                width=page_result.get('width', 0),
-                height=page_result.get('height', 0)
-            )
-
             # Apply gap filling if enabled and raw regions available
             if self.gap_filling_service and raw_text_regions:
                 # Filter raw regions for current page
@@ -556,9 +667,19 @@ class OCRToUnifiedConverter:
     def _convert_pp3_element(
         self,
         elem_data: Dict[str, Any],
-        page_idx: int
+        page_idx: int,
+        page_width: float = 0,
+        page_height: float = 0
     ) -> Optional[DocumentElement]:
-        """Convert PP-StructureV3 element to DocumentElement."""
+        """
+        Convert PP-StructureV3 element to DocumentElement.
+
+        Args:
+            elem_data: Element data from PP-StructureV3
+            page_idx: Page index (0-based)
+            page_width: Page width for coordinate validation
+            page_height: Page height for coordinate validation
+        """
         try:
             # Extract bbox
             bbox_data = elem_data.get('bbox', [0, 0, 0, 0])
@@ -597,18 +718,67 @@ class OCRToUnifiedConverter:
                 # Preserve cell_boxes and embedded_images in metadata for PDF generation
                 # These are extracted by PP-StructureV3 and provide accurate cell positioning
                 if 'cell_boxes' in elem_data:
-                    elem_data.setdefault('metadata', {})['cell_boxes'] = elem_data['cell_boxes']
+                    cell_boxes = elem_data['cell_boxes']
-                    elem_data['metadata']['cell_boxes_source'] = elem_data.get('cell_boxes_source', 'table_res_list')
+                    elem_data.setdefault('metadata', {})['cell_boxes_source'] = elem_data.get('cell_boxes_source', 'table_res_list')
+
+                    # Validate cell_boxes coordinates if page dimensions are available
+                    if page_width > 0 and page_height > 0:
+                        validation = validate_cell_boxes(
+                            cell_boxes=cell_boxes,
+                            table_bbox=bbox_data,
+                            page_width=page_width,
+                            page_height=page_height
+                        )
+
+                        if not validation['valid']:
+                            elem_data['metadata']['cell_boxes_validation'] = {
+                                'valid': False,
+                                'invalid_count': validation['invalid_count'],
+                                'total_count': len(cell_boxes),
+                                'needs_fallback': validation['needs_fallback']
+                            }
+                            # Use clamped boxes instead of invalid ones
+                            elem_data['metadata']['cell_boxes'] = validation['clamped_boxes']
+                            elem_data['metadata']['cell_boxes_original'] = cell_boxes
+
+                            if validation['needs_fallback']:
+                                logger.warning(
+                                    f"Table {elem_data.get('element_id')}: "
+                                    f"{validation['invalid_count']}/{len(cell_boxes)} cell_boxes invalid, "
+                                    f"fallback recommended"
+                                )
+                        else:
+                            elem_data['metadata']['cell_boxes'] = cell_boxes
+                            elem_data['metadata']['cell_boxes_validation'] = {'valid': True}
+                    else:
+                        # No page dimensions available, store as-is
+                        elem_data['metadata']['cell_boxes'] = cell_boxes
+
                 if 'embedded_images' in elem_data:
                     elem_data.setdefault('metadata', {})['embedded_images'] = elem_data['embedded_images']
-            elif element_type in [ElementType.IMAGE, ElementType.FIGURE]:
+            elif element_type in [
-                # For images, use metadata dict as content
+                ElementType.IMAGE, ElementType.FIGURE, ElementType.CHART,
+                ElementType.DIAGRAM, ElementType.LOGO, ElementType.STAMP
+            ]:
+                # For all visual elements, use metadata dict as content
+                # Priority: saved_path > img_path (PP-StructureV3 uses saved_path)
+                image_path = (
+                    elem_data.get('saved_path') or
+                    elem_data.get('img_path') or
+                    ''
+                )
                 content = {
-                    'path': elem_data.get('img_path', ''),
+                    'saved_path': image_path,  # Preserve original path key
+                    'path': image_path,        # For backward compatibility
                     'width': elem_data.get('width', 0),
                     'height': elem_data.get('height', 0),
                     'format': elem_data.get('format', 'unknown')
                 }
+                if not image_path:
+                    logger.warning(
+                        f"Visual element {element_type.value} missing image path: "
+                        f"saved_path={elem_data.get('saved_path')}, img_path={elem_data.get('img_path')}"
+                    )
             else:
                 content = elem_data.get('content', '')
 
@@ -1139,10 +1309,18 @@ class OCRToUnifiedConverter:
         for page_idx, page_data in enumerate(pages_data):
             elements = []
 
+            # Get page dimensions first
+            page_width = page_data.get('width', 0)
+            page_height = page_data.get('height', 0)
+
             # Process each element in the page
             if 'elements' in page_data:
                 for elem_data in page_data['elements']:
-                    element = self._convert_pp3_element(elem_data, page_idx)
+                    element = self._convert_pp3_element(
+                        elem_data, page_idx,
+                        page_width=page_width,
+                        page_height=page_height
+                    )
                     if element:
                         elements.append(element)
 
@@ -1150,8 +1328,8 @@ class OCRToUnifiedConverter:
             page = Page(
                 page_number=page_idx + 1,
                 dimensions=Dimensions(
-                    width=page_data.get('width', 0),
+                    width=page_width,
-                    height=page_data.get('height', 0)
+                    height=page_height
                 ),
                 elements=elements,
                 metadata={'reading_order': self._calculate_reading_order(elements)}

backend/app/services/pdf_generator_service.py

@@ -3371,18 +3371,21 @@ class PDFGeneratorService:
             "rows": 6,
             "cols": 2,
             "cells": [
-                {"row": 0, "col": 0, "content": "..."},
+                {"row": 0, "col": 0, "content": "...", "row_span": 1, "col_span": 2},
                 {"row": 0, "col": 1, "content": "..."},
                 ...
             ]
         }
 
-        Returns format compatible with HTMLTableParser output:
+        Returns format compatible with HTMLTableParser output (with colspan/rowspan/col):
         [
-            {"cells": [{"text": "..."}, {"text": "..."}]},  # row 0
+            {"cells": [{"text": "...", "colspan": 1, "rowspan": 1, "col": 0}, ...]},
-            {"cells": [{"text": "..."}, {"text": "..."}]},  # row 1
+            {"cells": [{"text": "...", "colspan": 1, "rowspan": 1, "col": 0}, ...]},
             ...
         ]
+
+        Note: This returns actual cells per row with their absolute column positions.
+        The table renderer uses 'col' to place cells correctly in the grid.
         """
         try:
             num_rows = content.get('rows', 0)
@@ -3392,21 +3395,39 @@ class PDFGeneratorService:
             if not cells or num_rows == 0 or num_cols == 0:
                 return []
 
-            # Initialize rows structure
+            # Group cells by row
-            rows_data = []
+            cells_by_row = {}
-            for _ in range(num_rows):
-                rows_data.append({'cells': [{'text': ''} for _ in range(num_cols)]})
-
-            # Fill in cell content
             for cell in cells:
                 row_idx = cell.get('row', 0)
-                col_idx = cell.get('col', 0)
+                if row_idx not in cells_by_row:
+                    cells_by_row[row_idx] = []
+                cells_by_row[row_idx].append(cell)
+
+            # Sort cells within each row by column
+            for row_idx in cells_by_row:
+                cells_by_row[row_idx].sort(key=lambda c: c.get('col', 0))
+
+            # Build rows structure with colspan/rowspan info and absolute col position
+            rows_data = []
+            for row_idx in range(num_rows):
+                row_cells = []
+                if row_idx in cells_by_row:
+                    for cell in cells_by_row[row_idx]:
                         cell_content = cell.get('content', '')
+                        row_span = cell.get('row_span', 1) or 1
+                        col_span = cell.get('col_span', 1) or 1
+                        col_idx = cell.get('col', 0)
 
-                if 0 <= row_idx < num_rows and 0 <= col_idx < num_cols:
+                        row_cells.append({
-                    rows_data[row_idx]['cells'][col_idx]['text'] = str(cell_content) if cell_content else ''
+                            'text': str(cell_content) if cell_content else '',
+                            'rowspan': row_span,
+                            'colspan': col_span,
+                            'col': col_idx  # Absolute column position
+                        })
 
-            logger.debug(f"Built {num_rows} rows from cells dict")
+                rows_data.append({'cells': row_cells})
+
+            logger.debug(f"Built {num_rows} rows from cells dict with span info")
             return rows_data
 
         except Exception as e:
@@ -3471,19 +3492,115 @@ class PDFGeneratorService:
             table_width = bbox.x1 - bbox.x0
             table_height = bbox.y1 - bbox.y0
 
-            # Build table data for ReportLab
-            table_content = []
-            for row in rows:
-                row_data = [cell['text'].strip() for cell in row['cells']]
-                table_content.append(row_data)
-
             # Create table
             from reportlab.platypus import Table, TableStyle
             from reportlab.lib import colors
 
-            # Determine number of rows and columns for cell_boxes calculation
+            # Determine grid size from rows structure
+            # Note: rows may have 'col' attribute for absolute positioning (from Direct extraction)
+            # or may be sequential (from HTML parsing)
             num_rows = len(rows)
-            max_cols = max(len(row['cells']) for row in rows) if rows else 0
+
+            # Check if cells have absolute column positions
+            has_absolute_cols = any(
+                'col' in cell
+                for row in rows
+                for cell in row['cells']
+            )
+
+            # Calculate actual number of columns
+            max_cols = 0
+            if has_absolute_cols:
+                # Use absolute col positions + colspan to find max column
+                for row in rows:
+                    for cell in row['cells']:
+                        col = cell.get('col', 0)
+                        colspan = cell.get('colspan', 1)
+                        max_cols = max(max_cols, col + colspan)
+            else:
+                # Sequential cells: sum up colspans
+                for row in rows:
+                    col_pos = 0
+                    for cell in row['cells']:
+                        colspan = cell.get('colspan', 1)
+                        col_pos += colspan
+                    max_cols = max(max_cols, col_pos)
+
+            # Build table data for ReportLab with proper grid structure
+            # ReportLab needs a full grid with placeholders for spanned cells
+            # and SPAN commands to merge them
+            table_content = []
+            span_commands = []
+            covered = set()  # Track cells covered by spans
+
+            # First pass: mark covered cells and collect SPAN commands
+            for row_idx, row in enumerate(rows):
+                if has_absolute_cols:
+                    # Use absolute column positions
+                    for cell in row['cells']:
+                        col_pos = cell.get('col', 0)
+                        colspan = cell.get('colspan', 1)
+                        rowspan = cell.get('rowspan', 1)
+
+                        # Mark cells covered by this span
+                        if colspan > 1 or rowspan > 1:
+                            for r in range(row_idx, row_idx + rowspan):
+                                for c in range(col_pos, col_pos + colspan):
+                                    if (r, c) != (row_idx, col_pos):
+                                        covered.add((r, c))
+                            # Add SPAN command for ReportLab
+                            span_commands.append((
+                                'SPAN',
+                                (col_pos, row_idx),
+                                (col_pos + colspan - 1, row_idx + rowspan - 1)
+                            ))
+                else:
+                    # Sequential positioning
+                    col_pos = 0
+                    for cell in row['cells']:
+                        while (row_idx, col_pos) in covered:
+                            col_pos += 1
+
+                        colspan = cell.get('colspan', 1)
+                        rowspan = cell.get('rowspan', 1)
+
+                        if colspan > 1 or rowspan > 1:
+                            for r in range(row_idx, row_idx + rowspan):
+                                for c in range(col_pos, col_pos + colspan):
+                                    if (r, c) != (row_idx, col_pos):
+                                        covered.add((r, c))
+                            span_commands.append((
+                                'SPAN',
+                                (col_pos, row_idx),
+                                (col_pos + colspan - 1, row_idx + rowspan - 1)
+                            ))
+                        col_pos += colspan
+
+            # Second pass: build content grid
+            for row_idx in range(num_rows):
+                row_data = [''] * max_cols
+
+                if row_idx < len(rows):
+                    if has_absolute_cols:
+                        # Place cells at their absolute positions
+                        for cell in rows[row_idx]['cells']:
+                            col_pos = cell.get('col', 0)
+                            if col_pos < max_cols:
+                                row_data[col_pos] = cell['text'].strip()
+                    else:
+                        # Sequential placement
+                        col_pos = 0
+                        for cell in rows[row_idx]['cells']:
+                            while col_pos < max_cols and (row_idx, col_pos) in covered:
+                                col_pos += 1
+                            if col_pos < max_cols:
+                                row_data[col_pos] = cell['text'].strip()
+                                colspan = cell.get('colspan', 1)
+                                col_pos += colspan
+
+                table_content.append(row_data)
+
+            logger.debug(f"Built table grid: {num_rows} rows × {max_cols} cols, {len(span_commands)} span commands (absolute_cols={has_absolute_cols})")
 
             # Use original column widths from extraction if available
             # Otherwise try to compute from cell_boxes (from PP-StructureV3)
@@ -3517,7 +3634,7 @@ class PDFGeneratorService:
             # Apply style with minimal padding to reduce table extension
             # Use Chinese font to support special characters (℃, μm, ≦, ×, Ω, etc.)
             font_for_table = self.font_name if self.font_registered else 'Helvetica'
-            style = TableStyle([
+            style_commands = [
                 ('GRID', (0, 0), (-1, -1), 0.5, colors.grey),
                 ('FONTNAME', (0, 0), (-1, -1), font_for_table),
                 ('FONTSIZE', (0, 0), (-1, -1), 8),
@@ -3529,7 +3646,13 @@ class PDFGeneratorService:
                 ('BOTTOMPADDING', (0, 0), (-1, -1), 0),
                 ('LEFTPADDING', (0, 0), (-1, -1), 1),
                 ('RIGHTPADDING', (0, 0), (-1, -1), 1),
-            ])
+            ]
+            # Add span commands for merged cells
+            style_commands.extend(span_commands)
+            if span_commands:
+                logger.info(f"Applied {len(span_commands)} SPAN commands for merged cells")
+
+            style = TableStyle(style_commands)
             t.setStyle(style)
 
             # Use canvas scaling as fallback to fit table within bbox
@@ -4350,30 +4473,97 @@ class PDFGeneratorService:
                         # Replace newlines with <br/>
                         safe_content = safe_content.replace('\n', '<br/>')
 
-                        # Calculate font size from bbox height, but keep minimum 10pt
+                        # Get original font size from style info
-                        font_size = max(box_height * 0.7, 10)
+                        style_info = elem.get('style', {})
-                        font_size = min(font_size, 24)  # Cap at 24pt
+                        original_font_size = style_info.get('font_size', 12.0)
 
-                        # Create style for this element
+                        # Detect vertical text (Y-axis labels, etc.)
-                        elem_style = ParagraphStyle(
+                        # Vertical text has aspect_ratio (height/width) > 2 and multiple characters
-                            f'elem_{id(elem)}',
+                        is_vertical_text = (
-                            parent=base_style,
+                            box_height > box_width * 2 and
-                            fontSize=font_size,
+                            len(content.strip()) > 1
-                            leading=font_size * 1.2,
+                        )
+
+                        if is_vertical_text:
+                            # For vertical text, use original font size and rotate
+                            font_size = min(original_font_size, box_width * 0.9)
+                            font_size = max(font_size, 6)  # Minimum 6pt
+
+                            # Save canvas state for rotation
+                            pdf_canvas.saveState()
+
+                            # Convert to PDF coordinates
+                            pdf_y_center = current_page_height - (y0 + y1) / 2
+                            x_center = (x0 + x1) / 2
+
+                            # Translate to center, rotate, translate back
+                            pdf_canvas.translate(x_center, pdf_y_center)
+                            pdf_canvas.rotate(90)
+
+                            # Set font and draw text centered
+                            pdf_canvas.setFont(
+                                self.font_name if self.font_registered else 'Helvetica',
+                                font_size
+                            )
+                            # Draw text at origin (since we translated to center)
+                            text_width = pdf_canvas.stringWidth(
+                                safe_content.replace('&amp;', '&').replace('&lt;', '<').replace('&gt;', '>'),
+                                self.font_name if self.font_registered else 'Helvetica',
+                                font_size
+                            )
+                            pdf_canvas.drawString(-text_width / 2, -font_size / 3,
+                                safe_content.replace('&amp;', '&').replace('&lt;', '<').replace('&gt;', '>'))
+
+                            pdf_canvas.restoreState()
+                        else:
+                            # For horizontal text, dynamically fit text within bbox
+                            # Start with original font size and reduce until text fits
+                            MIN_FONT_SIZE = 6
+                            MAX_FONT_SIZE = 14
+
+                            if original_font_size > 0:
+                                start_font_size = min(original_font_size, MAX_FONT_SIZE)
+                            else:
+                                start_font_size = min(box_height * 0.7, MAX_FONT_SIZE)
+
+                            font_size = max(start_font_size, MIN_FONT_SIZE)
+
+                            # Try progressively smaller font sizes until text fits
+                            para = None
+                            para_height = box_height + 1  # Start with height > box to enter loop
+
+                            while font_size >= MIN_FONT_SIZE and para_height > box_height:
+                                elem_style = ParagraphStyle(
+                                    f'elem_{id(elem)}_{font_size}',
+                                    parent=base_style,
+                                    fontSize=font_size,
+                                    leading=font_size * 1.15,  # Tighter leading
                                 )
 
-                        # Create paragraph
                                 para = Paragraph(safe_content, elem_style)
+                                para_width, para_height = para.wrap(box_width, box_height * 3)
 
-                        # Calculate available width and height
+                                if para_height <= box_height:
-                        available_width = box_width
+                                    break  # Text fits!
-                        available_height = box_height * 2  # Allow overflow
 
-                        # Wrap the paragraph
+                                font_size -= 0.5  # Reduce font size and try again
-                        para_width, para_height = para.wrap(available_width, available_height)
+
+                            # Ensure minimum font size
+                            if font_size < MIN_FONT_SIZE:
+                                font_size = MIN_FONT_SIZE
+                                elem_style = ParagraphStyle(
+                                    f'elem_{id(elem)}_min',
+                                    parent=base_style,
+                                    fontSize=font_size,
+                                    leading=font_size * 1.15,
+                                )
+                                para = Paragraph(safe_content, elem_style)
+                                para_width, para_height = para.wrap(box_width, box_height * 3)
 
                             # Convert to PDF coordinates (y from bottom)
-                        pdf_y = current_page_height - y0 - para_height
+                            # Clip to bbox height to prevent overflow
+                            actual_height = min(para_height, box_height)
+                            pdf_y = current_page_height - y0 - actual_height
 
                             # Draw the paragraph
                             para.drawOn(pdf_canvas, x0, pdf_y)
@@ -4451,13 +4641,47 @@ class PDFGeneratorService:
             pdf_y_bottom = page_height - ty1
             pdf_canvas.rect(tx0, pdf_y_bottom, table_width, table_height, stroke=1, fill=0)
 
-            # Step 2: Draw cell borders using cell_boxes
+            # Step 2: Get or calculate cell boxes
             cell_boxes = metadata.get('cell_boxes', [])
-            if cell_boxes:
+
+            # If no cell_boxes, calculate from column_widths and row_heights
+            if not cell_boxes:
+                column_widths = metadata.get('column_widths', [])
+                row_heights = metadata.get('row_heights', [])
+
+                if column_widths and row_heights:
+                    # Calculate cell positions from widths and heights
+                    cell_boxes = []
+                    rows = content.get('rows', len(row_heights)) if isinstance(content, dict) else len(row_heights)
+                    cols = content.get('cols', len(column_widths)) if isinstance(content, dict) else len(column_widths)
+
+                    # Calculate cumulative positions
+                    x_positions = [tx0]
+                    for w in column_widths[:cols]:
+                        x_positions.append(x_positions[-1] + w)
+
+                    y_positions = [ty0]
+                    for h in row_heights[:rows]:
+                        y_positions.append(y_positions[-1] + h)
+
+                    # Create cell boxes for each cell (row-major order)
+                    for row_idx in range(rows):
+                        for col_idx in range(cols):
+                            if col_idx < len(x_positions) - 1 and row_idx < len(y_positions) - 1:
+                                cx0 = x_positions[col_idx]
+                                cy0 = y_positions[row_idx]
+                                cx1 = x_positions[col_idx + 1]
+                                cy1 = y_positions[row_idx + 1]
+                                cell_boxes.append([cx0, cy0, cx1, cy1])
+
+                    logger.debug(f"Calculated {len(cell_boxes)} cell boxes from {cols} cols x {rows} rows")
+
             # Normalize cell boxes for grid alignment
-                if hasattr(self, '_normalize_cell_boxes_to_grid'):
+            if cell_boxes and hasattr(self, '_normalize_cell_boxes_to_grid'):
                 cell_boxes = self._normalize_cell_boxes_to_grid(cell_boxes)
 
+            # Draw cell borders
+            if cell_boxes:
                 pdf_canvas.setLineWidth(0.5)
                 for box in cell_boxes:
                     if len(box) >= 4:

backend/app/services/pp_structure_enhanced.py

@@ -558,8 +558,8 @@ class PPStructureEnhanced:
                         element['embedded_images'] = embedded_images
                         logger.info(f"[TABLE] Embedded {len(embedded_images)} images into table")
 
-            # Special handling for images/figures/stamps (visual elements that need cropping)
+            # Special handling for images/figures/charts/stamps (visual elements that need cropping)
-            elif mapped_type in [ElementType.IMAGE, ElementType.FIGURE, ElementType.STAMP, ElementType.LOGO]:
+            elif mapped_type in [ElementType.IMAGE, ElementType.FIGURE, ElementType.CHART, ElementType.DIAGRAM, ElementType.STAMP, ElementType.LOGO]:
                 # Save image if path provided
                 if 'img_path' in item and output_dir:
                     saved_path = self._save_image(item['img_path'], output_dir, element['element_id'])

backend/tests/debug_table_cells.py

@@ -0,0 +1,43 @@
+"""Debug PyMuPDF table.cells structure"""
+import sys
+from pathlib import Path
+sys.path.insert(0, str(Path(__file__).parent.parent))
+
+import fitz
+
+pdf_path = Path(__file__).parent.parent.parent / "demo_docs" / "edit3.pdf"
+doc = fitz.open(str(pdf_path))
+page = doc[0]
+
+tables = page.find_tables()
+for idx, table in enumerate(tables.tables):
+    data = table.extract()
+    num_rows = len(data)
+    num_cols = max(len(row) for row in data) if data else 0
+
+    print(f"Table {idx}:")
+    print(f"  table.extract() dimensions: {num_rows} rows x {num_cols} cols")
+    print(f"  Expected positions: {num_rows * num_cols}")
+
+    cell_rects = getattr(table, 'cells', None)
+    if cell_rects:
+        print(f"  table.cells length: {len(cell_rects)}")
+        none_count = sum(1 for c in cell_rects if c is None)
+        actual_count = sum(1 for c in cell_rects if c is not None)
+        print(f"  None cells: {none_count}")
+        print(f"  Actual cells: {actual_count}")
+
+        # Check if cell_rects matches grid size
+        if len(cell_rects) != num_rows * num_cols:
+            print(f"  WARNING: cell_rects length ({len(cell_rects)}) != grid size ({num_rows * num_cols})")
+
+        # Show first few cells
+        print(f"  First 5 cells: {cell_rects[:5]}")
+    else:
+        print(f"  table.cells: NOT AVAILABLE")
+
+    # Check row_count and col_count
+    print(f"  table.row_count: {getattr(table, 'row_count', 'N/A')}")
+    print(f"  table.col_count: {getattr(table, 'col_count', 'N/A')}")
+
+doc.close()

backend/tests/debug_table_cells2.py

@@ -0,0 +1,48 @@
+"""Debug PyMuPDF table structure - find merge info"""
+import sys
+from pathlib import Path
+sys.path.insert(0, str(Path(__file__).parent.parent))
+
+import fitz
+
+pdf_path = Path(__file__).parent.parent.parent / "demo_docs" / "edit3.pdf"
+doc = fitz.open(str(pdf_path))
+page = doc[0]
+
+tables = page.find_tables()
+for idx, table in enumerate(tables.tables):
+    print(f"\nTable {idx}:")
+
+    # Check all available attributes
+    print(f"  Available attributes: {[a for a in dir(table) if not a.startswith('_')]}")
+
+    # Try to get header info
+    if hasattr(table, 'header'):
+        print(f"  header: {table.header}")
+
+    # Check for cells info
+    cell_rects = table.cells
+    print(f"  cells count: {len(cell_rects)}")
+
+    # Get the extracted data
+    data = table.extract()
+    print(f"  extract() shape: {len(data)} x {max(len(r) for r in data)}")
+
+    # Check if there's a way to map cells to grid positions
+    # Look at the pandas output which might have merge info
+    try:
+        df = table.to_pandas()
+        print(f"  pandas shape: {df.shape}")
+    except Exception as e:
+        print(f"  pandas error: {e}")
+
+    # Check the TableRow objects if available
+    if hasattr(table, 'rows'):
+        rows = table.rows
+        print(f"  rows: {len(rows)}")
+        for ri, row in enumerate(rows[:3]):  # first 3 rows
+            print(f"    row {ri}: {len(row.cells)} cells")
+            for ci, cell in enumerate(row.cells[:5]):  # first 5 cells
+                print(f"      cell {ci}: bbox={cell}")
+
+doc.close()

backend/tests/generate_test_pdf.py

@@ -0,0 +1,111 @@
+"""
+Generate test PDF to verify Phase 1 fixes
+"""
+
+import sys
+import os
+from pathlib import Path
+
+# Add backend to path
+sys.path.insert(0, str(Path(__file__).parent.parent))
+
+from app.services.direct_extraction_engine import DirectExtractionEngine
+from app.services.pdf_generator_service import PDFGeneratorService
+from app.services.unified_document_exporter import UnifiedDocumentExporter
+
+
+def generate_test_pdf(input_pdf: str, output_dir: Path):
+    """Generate test PDF using Direct Track extraction"""
+
+    input_path = Path(input_pdf)
+    output_dir.mkdir(parents=True, exist_ok=True)
+
+    print(f"Processing: {input_path.name}")
+    print(f"Output dir: {output_dir}")
+
+    # Step 1: Extract with Direct Track
+    engine = DirectExtractionEngine(
+        enable_table_detection=True,
+        enable_image_extraction=True,
+        min_image_area=200.0,  # Filter tiny images
+        enable_whiteout_detection=True,
+        enable_content_sanitization=True
+    )
+
+    unified_doc = engine.extract(input_path, output_dir=output_dir)
+
+    # Print extraction stats
+    print(f"\n=== Extraction Results ===")
+    print(f"Document ID: {unified_doc.document_id}")
+    print(f"Pages: {len(unified_doc.pages)}")
+
+    table_count = 0
+    image_count = 0
+    merged_cells = 0
+    total_cells = 0
+
+    for page in unified_doc.pages:
+        for elem in page.elements:
+            if elem.type.value == 'table':
+                table_count += 1
+                if elem.content and hasattr(elem.content, 'cells'):
+                    total_cells += len(elem.content.cells)
+                    for cell in elem.content.cells:
+                        if cell.row_span > 1 or cell.col_span > 1:
+                            merged_cells += 1
+            elif elem.type.value == 'image':
+                image_count += 1
+
+    print(f"Tables: {table_count}")
+    print(f"  - Total cells: {total_cells}")
+    print(f"  - Merged cells: {merged_cells}")
+    print(f"Images: {image_count}")
+
+    # Step 2: Export to JSON
+    exporter = UnifiedDocumentExporter()
+    json_path = output_dir / f"{input_path.stem}_result.json"
+    exporter.export_to_json(unified_doc, json_path)
+    print(f"\nJSON saved: {json_path}")
+
+    # Step 3: Generate layout PDF
+    pdf_generator = PDFGeneratorService()
+    pdf_path = output_dir / f"{input_path.stem}_layout.pdf"
+
+    try:
+        pdf_generator.generate_from_unified_document(
+            unified_doc=unified_doc,
+            output_path=pdf_path,
+            source_file_path=input_path
+        )
+        print(f"PDF saved: {pdf_path}")
+        return pdf_path
+    except Exception as e:
+        print(f"PDF generation error: {e}")
+        import traceback
+        traceback.print_exc()
+        return None
+
+
+if __name__ == "__main__":
+    # Test with edit3.pdf (has complex tables with merging)
+    demo_docs = Path(__file__).parent.parent.parent / "demo_docs"
+    output_base = Path(__file__).parent.parent / "storage" / "test_phase1"
+
+    # Process edit3.pdf
+    edit3_pdf = demo_docs / "edit3.pdf"
+    if edit3_pdf.exists():
+        output_dir = output_base / "edit3"
+        result = generate_test_pdf(str(edit3_pdf), output_dir)
+        if result:
+            print(f"\n✓ Test PDF generated: {result}")
+
+    # Also process edit.pdf for comparison
+    edit_pdf = demo_docs / "edit.pdf"
+    if edit_pdf.exists():
+        output_dir = output_base / "edit"
+        result = generate_test_pdf(str(edit_pdf), output_dir)
+        if result:
+            print(f"\n✓ Test PDF generated: {result}")
+
+    print(f"\n=== Output Location ===")
+    print(f"{output_base}")

backend/tests/test_phase1_fixes.py

@@ -0,0 +1,285 @@
+"""
+Phase 1 Bug Fixes Verification Tests
+
+Tests for:
+1.1 Direct Track table cell merging
+1.2 OCR Track image path preservation
+1.3 Cell boxes coordinate validation
+1.4 Tiny decoration image filtering
+1.5 Covering image removal
+"""
+
+import sys
+import os
+from pathlib import Path
+
+# Add backend to path
+sys.path.insert(0, str(Path(__file__).parent.parent))
+
+import fitz
+from app.services.direct_extraction_engine import DirectExtractionEngine
+from app.services.ocr_to_unified_converter import validate_cell_boxes
+from app.models.unified_document import TableCell
+
+
+def test_1_1_table_cell_merging():
+    """Test 1.1.5: Verify edit3.pdf returns correct merged cells"""
+    print("\n" + "="*60)
+    print("TEST 1.1: Direct Track Table Cell Merging")
+    print("="*60)
+
+    pdf_path = Path(__file__).parent.parent.parent / "demo_docs" / "edit3.pdf"
+    if not pdf_path.exists():
+        print(f"SKIP: {pdf_path} not found")
+        return False
+
+    doc = fitz.open(str(pdf_path))
+
+    total_cells = 0
+    merged_cells = 0
+
+    for page_num, page in enumerate(doc):
+        tables = page.find_tables()
+        for table_idx, table in enumerate(tables.tables):
+            data = table.extract()
+            cell_rects = getattr(table, 'cells', None)
+
+            if cell_rects:
+                num_rows = len(data)
+                num_cols = max(len(row) for row in data) if data else 0
+
+                # Count actual cells (non-None)
+                actual_cells = sum(1 for c in cell_rects if c is not None)
+                none_cells = sum(1 for c in cell_rects if c is None)
+
+                print(f"  Page {page_num}, Table {table_idx}:")
+                print(f"    Grid size: {num_rows} x {num_cols} = {num_rows * num_cols} positions")
+                print(f"    Actual cells: {actual_cells}")
+                print(f"    Merged positions (None): {none_cells}")
+
+                total_cells += actual_cells
+                if none_cells > 0:
+                    merged_cells += 1
+
+    doc.close()
+
+    print(f"\n  Total actual cells across all tables: {total_cells}")
+    print(f"  Tables with merging: {merged_cells}")
+
+    # According to PLAN.md, edit3.pdf should have 83 cells (not 204)
+    # The presence of None values indicates merging is detected
+    if total_cells > 0 and total_cells < 204:
+        print("  RESULT: PASS - Cell merging detected correctly")
+        return True
+    elif total_cells == 204:
+        print("  RESULT: FAIL - All cells treated as 1x1 (no merging detected)")
+        return False
+    else:
+        print(f"  RESULT: INCONCLUSIVE - {total_cells} cells found")
+        return None
+
+
+def test_1_3_cell_boxes_validation():
+    """Test 1.3: Verify cell_boxes coordinate validation"""
+    print("\n" + "="*60)
+    print("TEST 1.3: Cell Boxes Coordinate Validation")
+    print("="*60)
+
+    # Test case 1: Valid coordinates
+    valid_boxes = [
+        [10, 10, 100, 50],
+        [100, 10, 200, 50],
+        [10, 50, 200, 100]
+    ]
+    result = validate_cell_boxes(valid_boxes, [0, 0, 300, 200], 300, 200)
+    print(f"  Valid boxes: valid={result['valid']}, invalid_count={result['invalid_count']}")
+    assert result['valid'], "Valid boxes should pass validation"
+
+    # Test case 2: Out of bounds coordinates
+    invalid_boxes = [
+        [-10, 10, 100, 50],    # x0 < 0
+        [10, 10, 400, 50],     # x1 > page_width
+        [10, 10, 100, 300]     # y1 > page_height
+    ]
+    result = validate_cell_boxes(invalid_boxes, [0, 0, 300, 200], 300, 200)
+    print(f"  Invalid boxes: valid={result['valid']}, invalid_count={result['invalid_count']}")
+    assert not result['valid'], "Invalid boxes should fail validation"
+    assert result['invalid_count'] == 3, "Should detect 3 invalid boxes"
+
+    # Test case 3: Clamping
+    assert len(result['clamped_boxes']) == 3, "Should return clamped boxes"
+    clamped = result['clamped_boxes'][0]
+    assert clamped[0] >= 0, "Clamped x0 should be >= 0"
+
+    print("  RESULT: PASS - Coordinate validation works correctly")
+    return True
+
+
+def test_1_4_tiny_image_filtering():
+    """Test 1.4: Verify tiny decoration image filtering"""
+    print("\n" + "="*60)
+    print("TEST 1.4: Tiny Decoration Image Filtering")
+    print("="*60)
+
+    pdf_path = Path(__file__).parent.parent.parent / "demo_docs" / "edit3.pdf"
+    if not pdf_path.exists():
+        print(f"SKIP: {pdf_path} not found")
+        return None
+
+    doc = fitz.open(str(pdf_path))
+
+    tiny_count = 0
+    normal_count = 0
+    min_area = 200  # Same threshold as in DirectExtractionEngine
+
+    for page_num, page in enumerate(doc):
+        images = page.get_images()
+        for img in images:
+            xref = img[0]
+            rects = page.get_image_rects(xref)
+            if rects:
+                rect = rects[0]
+                area = (rect.x1 - rect.x0) * (rect.y1 - rect.y0)
+                if area < min_area:
+                    tiny_count += 1
+                    print(f"  Page {page_num}: Tiny image xref={xref}, area={area:.1f} px²")
+                else:
+                    normal_count += 1
+
+    doc.close()
+
+    print(f"\n  Tiny images (< {min_area} px²): {tiny_count}")
+    print(f"  Normal images: {normal_count}")
+
+    if tiny_count > 0:
+        print("  RESULT: PASS - Tiny images detected, will be filtered")
+        return True
+    else:
+        print("  RESULT: INFO - No tiny images found in test file")
+        return None
+
+
+def test_1_5_covering_image_detection():
+    """Test 1.5: Verify covering image detection"""
+    print("\n" + "="*60)
+    print("TEST 1.5: Covering Image Detection")
+    print("="*60)
+
+    pdf_path = Path(__file__).parent.parent.parent / "demo_docs" / "edit3.pdf"
+    if not pdf_path.exists():
+        print(f"SKIP: {pdf_path} not found")
+        return None
+
+    engine = DirectExtractionEngine(
+        enable_whiteout_detection=True,
+        whiteout_iou_threshold=0.8
+    )
+
+    doc = fitz.open(str(pdf_path))
+
+    total_covering = 0
+    for page_num, page in enumerate(doc):
+        result = engine._preprocess_page(page, page_num, doc)
+        covering_images = result.get('covering_images', [])
+
+        if covering_images:
+            print(f"  Page {page_num}: {len(covering_images)} covering images detected")
+            for img in covering_images[:3]:  # Show first 3
+                print(f"    - xref={img.get('xref')}, type={img.get('color_type')}, "
+                      f"bbox={[round(x, 1) for x in img.get('bbox', [])]}")
+            total_covering += len(covering_images)
+
+    doc.close()
+
+    print(f"\n  Total covering images detected: {total_covering}")
+
+    if total_covering > 0:
+        print("  RESULT: PASS - Covering images detected, will be filtered")
+        return True
+    else:
+        print("  RESULT: INFO - No covering images found in test file")
+        return None
+
+
+def test_direct_extraction_full():
+    """Full integration test for Direct Track extraction"""
+    print("\n" + "="*60)
+    print("INTEGRATION TEST: Direct Track Full Extraction")
+    print("="*60)
+
+    pdf_path = Path(__file__).parent.parent.parent / "demo_docs" / "edit3.pdf"
+    if not pdf_path.exists():
+        print(f"SKIP: {pdf_path} not found")
+        return None
+
+    engine = DirectExtractionEngine(
+        enable_table_detection=True,
+        enable_image_extraction=True,
+        min_image_area=200.0,
+        enable_whiteout_detection=True
+    )
+
+    try:
+        result = engine.extract(pdf_path)  # Pass Path object, not string
+
+        # Count elements
+        table_count = 0
+        image_count = 0
+        merged_table_count = 0
+
+        for page in result.pages:
+            for elem in page.elements:
+                if elem.type.value == 'table':
+                    table_count += 1
+                    if elem.content and hasattr(elem.content, 'cells'):
+                        # Check for merged cells
+                        for cell in elem.content.cells:
+                            if cell.row_span > 1 or cell.col_span > 1:
+                                merged_table_count += 1
+                                break
+                elif elem.type.value == 'image':
+                    image_count += 1
+
+        print(f"  Document ID: {result.document_id}")
+        print(f"  Pages: {len(result.pages)}")
+        print(f"  Tables: {table_count} (with merging: {merged_table_count})")
+        print(f"  Images: {image_count}")
+
+        print("  RESULT: PASS - Extraction completed successfully")
+        return True
+
+    except Exception as e:
+        print(f"  RESULT: FAIL - {e}")
+        import traceback
+        traceback.print_exc()
+        return False
+
+
+if __name__ == "__main__":
+    print("="*60)
+    print("Phase 1 Bug Fixes Verification Tests")
+    print("="*60)
+
+    results = {}
+
+    # Run tests
+    results['1.1_table_merging'] = test_1_1_table_cell_merging()
+    results['1.3_coord_validation'] = test_1_3_cell_boxes_validation()
+    results['1.4_tiny_filtering'] = test_1_4_tiny_image_filtering()
+    results['1.5_covering_detection'] = test_1_5_covering_image_detection()
+    results['integration'] = test_direct_extraction_full()
+
+    # Summary
+    print("\n" + "="*60)
+    print("TEST SUMMARY")
+    print("="*60)
+
+    for test_name, result in results.items():
+        status = "PASS" if result is True else "FAIL" if result is False else "SKIP/INFO"
+        print(f"  {test_name}: {status}")
+
+    passed = sum(1 for r in results.values() if r is True)
+    failed = sum(1 for r in results.values() if r is False)
+    skipped = sum(1 for r in results.values() if r is None)
+
+    print(f"\n  Total: {passed} passed, {failed} failed, {skipped} skipped/info")

frontend/src/components/ProcessingTrackSelector.tsx

@@ -0,0 +1,148 @@
+import { Card, CardContent, CardHeader, CardTitle } from '@/components/ui/card'
+import { Badge } from '@/components/ui/badge'
+import { Cpu, FileText, Sparkles, Info } from 'lucide-react'
+import type { ProcessingTrack, DocumentAnalysisResponse } from '@/types/apiV2'
+
+interface ProcessingTrackSelectorProps {
+  value: ProcessingTrack | null  // null means "use system recommendation"
+  onChange: (track: ProcessingTrack | null) => void
+  documentAnalysis?: DocumentAnalysisResponse | null
+  disabled?: boolean
+}
+
+export default function ProcessingTrackSelector({
+  value,
+  onChange,
+  documentAnalysis,
+  disabled = false,
+}: ProcessingTrackSelectorProps) {
+  const recommendedTrack = documentAnalysis?.recommended_track
+
+  const tracks = [
+    {
+      id: null as ProcessingTrack | null,
+      name: '自動選擇',
+      description: '根據文件類型自動選擇最佳處理方式',
+      icon: Sparkles,
+      color: 'text-purple-600',
+      bgColor: 'bg-purple-50',
+      borderColor: 'border-purple-200',
+      recommended: false,
+    },
+    {
+      id: 'direct' as ProcessingTrack,
+      name: '直接提取 (DIRECT)',
+      description: '從 PDF 中直接提取文字圖層，適用於可編輯 PDF',
+      icon: FileText,
+      color: 'text-blue-600',
+      bgColor: 'bg-blue-50',
+      borderColor: 'border-blue-200',
+      recommended: recommendedTrack === 'direct',
+    },
+    {
+      id: 'ocr' as ProcessingTrack,
+      name: 'OCR 識別',
+      description: '使用光學字元識別處理圖片或掃描文件',
+      icon: Cpu,
+      color: 'text-green-600',
+      bgColor: 'bg-green-50',
+      borderColor: 'border-green-200',
+      recommended: recommendedTrack === 'ocr',
+    },
+  ]
+
+  return (
+    <Card>
+      <CardHeader>
+        <div className="flex items-center gap-3">
+          <div className="p-2 bg-primary/10 rounded-lg">
+            <Sparkles className="w-5 h-5 text-primary" />
+          </div>
+          <div>
+            <CardTitle>處理方式選擇</CardTitle>
+            <p className="text-sm text-muted-foreground mt-1">
+              選擇文件的處理方式，或讓系統自動判斷
+            </p>
+          </div>
+        </div>
+      </CardHeader>
+      <CardContent className="space-y-3">
+        {/* Info about override */}
+        {value !== null && recommendedTrack && value !== recommendedTrack && (
+          <div className="flex items-start gap-2 p-3 bg-amber-50 border border-amber-200 rounded-lg">
+            <Info className="w-4 h-4 text-amber-600 flex-shrink-0 mt-0.5" />
+            <p className="text-sm text-amber-800">
+              您已覆蓋系統建議。系統原本建議使用「{recommendedTrack === 'direct' ? '直接提取' : 'OCR 識別'}」方式處理此文件。
+            </p>
+          </div>
+        )}
+
+        {/* Track options */}
+        <div className="grid gap-3">
+          {tracks.map((track) => {
+            const isSelected = value === track.id
+            const Icon = track.icon
+
+            return (
+              <button
+                key={track.id ?? 'auto'}
+                type="button"
+                disabled={disabled}
+                onClick={() => onChange(track.id)}
+                className={`
+                  w-full p-4 rounded-lg border-2 text-left transition-all
+                  ${isSelected
+                    ? `${track.borderColor} ${track.bgColor}`
+                    : 'border-border hover:border-primary/30 hover:bg-muted/30'
+                  }
+                  ${disabled ? 'opacity-50 cursor-not-allowed' : 'cursor-pointer'}
+                `}
+              >
+                <div className="flex items-start gap-3">
+                  <div className={`p-2 rounded-lg ${isSelected ? track.bgColor : 'bg-muted'}`}>
+                    <Icon className={`w-5 h-5 ${isSelected ? track.color : 'text-muted-foreground'}`} />
+                  </div>
+                  <div className="flex-1 min-w-0">
+                    <div className="flex items-center gap-2">
+                      <span className={`font-medium ${isSelected ? track.color : ''}`}>
+                        {track.name}
+                      </span>
+                      {track.recommended && (
+                        <Badge variant="outline" className="text-xs bg-white">
+                          系統建議
+                        </Badge>
+                      )}
+                      {isSelected && (
+                        <Badge variant="default" className="text-xs">
+                          已選擇
+                        </Badge>
+                      )}
+                    </div>
+                    <p className="text-sm text-muted-foreground mt-1">
+                      {track.description}
+                    </p>
+                  </div>
+                </div>
+              </button>
+            )
+          })}
+        </div>
+
+        {/* Current analysis info */}
+        {documentAnalysis && (
+          <div className="pt-3 border-t border-border">
+            <div className="flex flex-wrap gap-x-4 gap-y-1 text-xs text-muted-foreground">
+              <span>文件分析信心度: {(documentAnalysis.confidence * 100).toFixed(0)}%</span>
+              {documentAnalysis.page_count && (
+                <span>頁數: {documentAnalysis.page_count}</span>
+              )}
+              {documentAnalysis.text_coverage !== null && (
+                <span>文字覆蓋率: {(documentAnalysis.text_coverage * 100).toFixed(1)}%</span>
+              )}
+            </div>
+          </div>
+        )}
+      </CardContent>
+    </Card>
+  )
+}

frontend/src/pages/ProcessingPage.tsx

@@ -8,14 +8,15 @@ import { Button } from '@/components/ui/button'
 import { Badge } from '@/components/ui/badge'
 import { useToast } from '@/components/ui/toast'
 import { apiClientV2 } from '@/services/apiV2'
-import { Play, CheckCircle, FileText, AlertCircle, Clock, Activity, Loader2, Info } from 'lucide-react'
+import { Play, CheckCircle, FileText, AlertCircle, Clock, Activity, Loader2 } from 'lucide-react'
 import LayoutModelSelector from '@/components/LayoutModelSelector'
 import PreprocessingSettings from '@/components/PreprocessingSettings'
 import PreprocessingPreview from '@/components/PreprocessingPreview'
 import TableDetectionSelector from '@/components/TableDetectionSelector'
+import ProcessingTrackSelector from '@/components/ProcessingTrackSelector'
 import TaskNotFound from '@/components/TaskNotFound'
 import { useTaskValidation } from '@/hooks/useTaskValidation'
-import type { LayoutModel, ProcessingOptions, PreprocessingMode, PreprocessingConfig, TableDetectionConfig, DocumentAnalysisResponse } from '@/types/apiV2'
+import type { LayoutModel, ProcessingOptions, PreprocessingMode, PreprocessingConfig, TableDetectionConfig, ProcessingTrack } from '@/types/apiV2'
 
 export default function ProcessingPage() {
   const { t } = useTranslation()
@@ -56,6 +57,9 @@ export default function ProcessingPage() {
     enable_region_detection: true,
   })
 
+  // Processing track override state (null = use system recommendation)
+  const [forceTrack, setForceTrack] = useState<ProcessingTrack | null>(null)
+
   // Analyze document to determine if OCR is needed (only for pending tasks)
   const { data: documentAnalysis, isLoading: isAnalyzing } = useQuery({
     queryKey: ['documentAnalysis', taskId],
@@ -65,16 +69,23 @@ export default function ProcessingPage() {
   })
 
   // Determine if preprocessing options should be shown
-  // Only show for OCR track files (images and non-editable PDFs)
+  // Show OCR options when:
-  const needsOcrTrack = documentAnalysis?.recommended_track === 'ocr' ||
+  // 1. User explicitly selected OCR track
+  // 2. OR system recommends OCR/hybrid track (and user hasn't overridden to direct)
+  // 3. OR still analyzing (show by default)
+  const needsOcrTrack = forceTrack === 'ocr' ||
+    (forceTrack === null && (
+      documentAnalysis?.recommended_track === 'ocr' ||
       documentAnalysis?.recommended_track === 'hybrid' ||
-    !documentAnalysis // Show by default while analyzing
+      !documentAnalysis
+    ))
 
   // Start OCR processing
   const processOCRMutation = useMutation({
     mutationFn: () => {
       const options: ProcessingOptions = {
-        use_dual_track: true,
+        use_dual_track: forceTrack === null,  // Only use dual-track auto-detection if not forcing
+        force_track: forceTrack || undefined,  // Pass force_track if user selected one
         language: 'ch',
         layout_model: layoutModel,
         preprocessing_mode: preprocessingMode,
@@ -392,53 +403,14 @@ export default function ProcessingPage() {
             </div>
           )}
 
-          {/* Document Analysis Info */}
+          {/* Processing Track Selector - Always show after analysis */}
-          {documentAnalysis && (
+          {!isAnalyzing && (
-            <Card className={documentAnalysis.recommended_track === 'direct' ? 'border-blue-200 bg-blue-50' : 'border-green-200 bg-green-50'}>
+            <ProcessingTrackSelector
-              <CardContent className="pt-4">
+              value={forceTrack}
-                <div className="flex items-start gap-3">
+              onChange={setForceTrack}
-                  <Info className={`w-5 h-5 flex-shrink-0 mt-0.5 ${documentAnalysis.recommended_track === 'direct' ? 'text-blue-600' : 'text-green-600'}`} />
+              documentAnalysis={documentAnalysis}
-                  <div className="flex-1">
+              disabled={processOCRMutation.isPending}
-                    {documentAnalysis.recommended_track === 'direct' ? (
+            />
-                      <>
-                        <p className="text-sm font-medium text-blue-800">此文件為可編輯 PDF</p>
-                        <p className="text-sm text-blue-700 mt-1">
-                          系統偵測到此 PDF 包含文字圖層，將使用直接文字提取方式處理。
-                          版面偵測和影像前處理設定不適用於此類文件。
-                        </p>
-                      </>
-                    ) : (
-                      <>
-                        <p className="text-sm font-medium text-green-800">
-                          {documentAnalysis.is_editable ? '混合文件' : '掃描文件 / 影像'}
-                        </p>
-                        <p className="text-sm text-green-700 mt-1">
-                          {documentAnalysis.reason}
-                        </p>
-                      </>
-                    )}
-                    <div className="flex flex-wrap gap-4 mt-2 text-xs">
-                      <span className={documentAnalysis.recommended_track === 'direct' ? 'text-blue-600' : 'text-green-600'}>
-                        處理方式: {documentAnalysis.recommended_track === 'direct' ? '直接提取' : documentAnalysis.recommended_track === 'ocr' ? 'OCR 識別' : '混合處理'}
-                      </span>
-                      {documentAnalysis.page_count && (
-                        <span className={documentAnalysis.recommended_track === 'direct' ? 'text-blue-600' : 'text-green-600'}>
-                          頁數: {documentAnalysis.page_count}
-                        </span>
-                      )}
-                      {documentAnalysis.text_coverage !== null && (
-                        <span className={documentAnalysis.recommended_track === 'direct' ? 'text-blue-600' : 'text-green-600'}>
-                          文字覆蓋率: {(documentAnalysis.text_coverage * 100).toFixed(1)}%
-                        </span>
-                      )}
-                      <span className={documentAnalysis.recommended_track === 'direct' ? 'text-blue-600' : 'text-green-600'}>
-                        信心度: {(documentAnalysis.confidence * 100).toFixed(0)}%
-                      </span>
-                    </div>
-                  </div>
-                </div>
-              </CardContent>
-            </Card>
           )}
 
           {/* OCR Track Options - Only show when document needs OCR */}

openspec/changes/refactor-dual-track-architecture/design.md

@@ -0,0 +1,240 @@
+# Design: Refactor Dual-Track Architecture
+
+## Context
+
+Tool_OCR 是一個雙軌制文件處理系統，支援：
+- **Direct Track**: 從可編輯 PDF 直接提取結構化內容
+- **OCR Track**: 使用 PaddleOCR + PP-StructureV3 進行光學字符識別
+
+目前系統存在以下技術債務：
+- OCRService (2,326 行) 承擔過多職責
+- PDFGeneratorService (4,644 行) 是單體服務
+- 記憶體管理分散在多個組件中
+- 已知 bug 影響輸出品質
+
+## Goals / Non-Goals
+
+### Goals
+- 修復 PLAN.md 中列出的所有已知 bug
+- 將 OCRService 拆分為 < 800 行的可維護單元
+- 將 PDFGeneratorService 拆分為 < 2,000 行
+- 簡化記憶體管理配置
+- 提升前端狀態管理一致性
+
+### Non-Goals
+- 不改變現有 API 契約
+- 不引入新的外部依賴
+- 不改變資料庫 schema
+- 不改變使用者介面
+
+## Decisions
+
+### Decision 1: 使用 PyMuPDF find_tables() 取代自定義表格檢測
+
+**選擇**: 使用 PyMuPDF 內建的 `page.find_tables()` API
+
+**理由**:
+- PyMuPDF 的表格檢測能正確識別合併單元格
+- 返回的 `table.cells` 結構包含 span 資訊
+- 減少自定義代碼維護負擔
+
+**替代方案**:
+- 改進 `_detect_tables_by_position()` 算法
+  - 優點：不依賴外部 API 變更
+  - 缺點：複雜度高，難以處理所有邊界情況
+- 使用 Camelot 或 Tabula
+  - 優點：成熟的表格提取庫
+  - 缺點：引入新依賴，增加系統複雜度
+
+### Decision 2: 使用 Strategy Pattern 重構服務層
+
+**選擇**: 引入 ProcessingOrchestrator 使用策略模式
+
+```python
+class ProcessingPipeline(Protocol):
+    def process(self, file_path: str, options: ProcessingOptions) -> UnifiedDocument:
+        ...
+
+class DirectPipeline(ProcessingPipeline):
+    def __init__(self, extraction_engine: DirectExtractionEngine):
+        self.engine = extraction_engine
+
+    def process(self, file_path, options):
+        return self.engine.extract(file_path)
+
+class OCRPipeline(ProcessingPipeline):
+    def __init__(self, ocr_service: OCRService, preprocessor: LayoutPreprocessingService):
+        self.ocr = ocr_service
+        self.preprocessor = preprocessor
+
+    def process(self, file_path, options):
+        # Preprocessing + OCR + Conversion
+        ...
+
+class ProcessingOrchestrator:
+    def __init__(self, detector: DocumentTypeDetector, pipelines: dict[str, ProcessingPipeline]):
+        self.detector = detector
+        self.pipelines = pipelines
+
+    def process(self, file_path, options):
+        track = options.force_track or self.detector.detect(file_path).track
+        return self.pipelines[track].process(file_path, options)
+```
+
+**理由**:
+- 職責分離：檢測、處理、轉換各自獨立
+- 易於測試：可以單獨測試每個 Pipeline
+- 易於擴展：新增處理方式只需添加新 Pipeline
+
+**替代方案**:
+- 使用 Chain of Responsibility
+  - 優點：更靈活的處理鏈
+  - 缺點：對於二選一的場景過於複雜
+- 保持現狀，只做代碼整理
+  - 優點：風險最低
+  - 缺點：無法解決根本問題
+
+### Decision 3: 分層提取 PDF 生成邏輯
+
+**選擇**: 將 PDFGeneratorService 拆分為三個模組
+
+```
+PDFGeneratorService (主要編排)
+├── PDFTableRenderer (表格渲染)
+│   ├── HTMLTableParser (HTML 表格解析)
+│   └── CellRenderer (單元格渲染)
+├── PDFFontManager (字體管理)
+│   ├── FontLoader (字體載入)
+│   └── FontFallback (字體 fallback)
+└── PDFLayoutEngine (版面配置)
+```
+
+**理由**:
+- 單一職責：每個模組專注一件事
+- 可重用：FontManager 可被其他服務使用
+- 易於測試：表格渲染可獨立測試
+
+### Decision 4: 統一記憶體策略引擎
+
+**選擇**: 合併記憶體管理組件為單一 MemoryPolicyEngine
+
+```python
+class MemoryPolicyEngine:
+    """統一的記憶體策略引擎"""
+
+    def __init__(self, config: MemoryConfig):
+        self.config = config
+        self._semaphore = asyncio.Semaphore(config.max_concurrent_predictions)
+
+    @property
+    def gpu_usage_percent(self) -> float:
+        # 統一的 GPU 使用率查詢
+        ...
+
+    def check_availability(self) -> MemoryStatus:
+        # 返回 AVAILABLE, WARNING, CRITICAL, EMERGENCY
+        ...
+
+    async def acquire_prediction_slot(self):
+        # 統一的並發控制
+        ...
+
+    def cleanup_if_needed(self):
+        # 根據狀態自動清理
+        ...
+
+@dataclass
+class MemoryConfig:
+    warning_threshold: float = 0.80      # 80%
+    critical_threshold: float = 0.95     # 95%
+    max_concurrent_predictions: int = 2
+    model_idle_timeout: int = 300        # 5 minutes
+```
+
+**理由**:
+- 減少配置項：從 8+ 降到 4 個核心配置
+- 簡化依賴：服務只需依賴一個記憶體引擎
+- 統一行為：所有記憶體決策在同一處做出
+
+### Decision 5: 使用 Zustand 管理任務狀態
+
+**選擇**: 新增 TaskStore 統一管理任務狀態
+
+```typescript
+interface TaskState {
+  currentTaskId: string | null;
+  tasks: Record<string, TaskDetail>;
+  processingStatus: Record<string, ProcessingStatus>;
+}
+
+interface TaskActions {
+  setCurrentTask: (taskId: string) => void;
+  updateTask: (taskId: string, updates: Partial<TaskDetail>) => void;
+  updateProcessingStatus: (taskId: string, status: ProcessingStatus) => void;
+  clearTasks: () => void;
+}
+
+const useTaskStore = create<TaskState & TaskActions>()(
+  persist(
+    (set) => ({
+      currentTaskId: null,
+      tasks: {},
+      processingStatus: {},
+      // ... actions
+    }),
+    { name: 'task-storage' }
+  )
+);
+```
+
+**理由**:
+- 一致性：與現有 uploadStore、authStore 模式一致
+- 可追蹤：任務狀態變更集中管理
+- 持久化：刷新頁面後狀態保留
+
+## Risks / Trade-offs
+
+| 風險 | 影響 | 緩解措施 |
+|------|------|----------|
+| PyMuPDF find_tables() API 變更 | 中 | 封裝為獨立函數，易於替換 |
+| 服務重構導致處理邏輯錯誤 | 高 | 保留原有測試，逐步重構 |
+| 記憶體引擎改變導致 OOM | 高 | 使用相同閾值，僅改變代碼結構 |
+| 前端狀態遷移導致 bug | 中 | 逐頁遷移，完整測試每個頁面 |
+
+## Migration Plan
+
+### Step 1: Bug Fixes (可獨立部署)
+1. 實現 PyMuPDF find_tables() 整合
+2. 修復 OCR Track 圖片路徑
+3. 添加 cell_boxes 座標驗證
+4. 測試並部署
+
+### Step 2: Service Refactoring (可獨立部署)
+1. 提取 ProcessingOrchestrator
+2. 提取 TableRenderer 和 FontManager
+3. 更新 OCRService 使用新組件
+4. 測試並部署
+
+### Step 3: Memory Management (可獨立部署)
+1. 實現 MemoryPolicyEngine
+2. 逐步遷移服務使用新引擎
+3. 移除舊組件
+4. 測試並部署
+
+### Step 4: Frontend Improvements (可獨立部署)
+1. 新增 TaskStore
+2. 遷移 ProcessingPage
+3. 遷移 TaskDetailPage
+4. 合併類型定義
+5. 測試並部署
+
+### Rollback Plan
+- 每個 Step 獨立部署，問題時可回滾到上一個穩定版本
+- Bug fixes 優先，確保基本功能正確
+- 重構不改變外部行為，回滾影響最小
+
+## Open Questions
+
+1. **PyMuPDF find_tables() 的版本相容性**: 需確認目前使用的 PyMuPDF 版本是否支援此 API
+2. **前端狀態持久化範圍**: 是否所有任務都需要持久化，還是只保留當前會話？
+3. **記憶體閾值調整**: 現有閾值是否經過生產驗證，可以直接沿用？

openspec/changes/refactor-dual-track-architecture/proposal.md

@@ -0,0 +1,68 @@
+# Change: Refactor Dual-Track Architecture
+
+## Why
+
+目前雙軌制 OCR 系統存在多個已知問題和架構債務：
+
+1. **Direct Track 表格問題**: `_detect_tables_by_position()` 無法識別合併單元格，導致 edit3.pdf 產生 204 個錯誤拆分的 cells（應為 83 個）
+2. **OCR Track 圖片路徑丟失**: CHART/DIAGRAM 等視覺元素的 `saved_path` 在轉換時丟失，導致圖片未放回 PDF
+3. **OCR Track cell_boxes 座標錯亂**: PP-StructureV3 返回的 cell_boxes 超出頁面邊界
+4. **服務層過度複雜**: OCRService (2,326 行) 承擔過多職責，難以維護和測試
+5. **PDF 生成器過於龐大**: PDFGeneratorService (4,644 行) 是單體服務，難以擴展
+
+## What Changes
+
+### Phase 1: 修復已知 Bug（優先級：最高）
+
+- **Direct Track 表格修復**: 改用 PyMuPDF `find_tables()` API 取代 `_detect_tables_by_position()`
+- **OCR Track 圖片路徑修復**: 擴展 `_convert_pp3_element` 處理所有視覺元素類型 (IMAGE, FIGURE, CHART, DIAGRAM, LOGO, STAMP)
+- **Cell boxes 座標驗證**: 添加邊界檢查，超出範圍時使用 CV 線檢測 fallback
+- **過濾極小裝飾圖片**: 過濾 < 200 px² 的圖片
+- **移除覆蓋圖像**: 在渲染階段過濾與 covering_images 重疊的圖片
+
+### Phase 2: 服務層重構（優先級：高）
+
+- **拆分 OCRService**: 提取獨立的 `ProcessingOrchestrator` 負責流程編排
+- **建立 Pipeline 模式**: 使用組合模式取代目前的聚合模式
+- **提取 TableRenderer**: 從 PDFGeneratorService 提取表格渲染邏輯
+- **提取 FontManager**: 從 PDFGeneratorService 提取字體管理邏輯
+
+### Phase 3: 記憶體管理簡化（優先級：中）
+
+- **統一記憶體策略**: 合併 MemoryManager、MemoryGuard、各類 Semaphore 為單一策略引擎
+- **簡化配置**: 減少 8+ 個記憶體相關配置項到核心 3-4 項
+
+### Phase 4: 前端狀態管理改進（優先級：中）
+
+- **新增 TaskStore**: 使用 Zustand 管理任務狀態，取代分散的 useState
+- **合併類型定義**: 統一 api.ts 和 apiV2.ts 為單一類型定義檔案
+
+## Impact
+
+- Affected specs: `document-processing`
+- Affected code:
+  - `backend/app/services/direct_extraction_engine.py` (表格檢測)
+  - `backend/app/services/ocr_to_unified_converter.py` (元素轉換)
+  - `backend/app/services/ocr_service.py` (服務編排)
+  - `backend/app/services/pdf_generator_service.py` (PDF 生成)
+  - `backend/app/services/memory_manager.py` (記憶體管理)
+  - `frontend/src/store/` (狀態管理)
+  - `frontend/src/types/` (類型定義)
+
+## Risk Assessment
+
+| 風險 | 嚴重性 | 緩解措施 |
+|------|--------|----------|
+| 表格渲染回歸 | 高 | 使用 edit.pdf 和 edit3.pdf 作為回歸測試 |
+| 記憶體管理變更導致 OOM | 高 | 保留現有閾值，僅重構代碼結構 |
+| 服務重構導致處理失敗 | 中 | 逐步重構，每階段完整測試 |
+
+## Success Metrics
+
+| 指標 | 目前 | 目標 |
+|------|------|------|
+| edit3.pdf Direct Track cells | 204 (錯誤) | 83 (正確) |
+| OCR Track 圖片放回率 | 0% | 100% |
+| cell_boxes 座標正確率 | ~40% | 100% |
+| OCRService 行數 | 2,326 | < 800 |
+| PDFGeneratorService 行數 | 4,644 | < 2,000 |

openspec/changes/refactor-dual-track-architecture/specs/document-processing/spec.md

@@ -0,0 +1,151 @@
+# document-processing Specification Delta
+
+## ADDED Requirements
+
+### Requirement: Table Cell Merging Detection
+The system SHALL correctly detect and preserve merged cells (rowspan/colspan) when extracting tables from PDF documents.
+
+#### Scenario: Detect merged cells in Direct Track
+- **WHEN** extracting tables from an editable PDF using Direct Track
+- **THEN** the system SHALL use PyMuPDF find_tables() API
+- **AND** correctly identify cells with rowspan > 1 or colspan > 1
+- **AND** preserve merge information in UnifiedDocument table structure
+- **AND** skip placeholder cells that are covered by merged cells
+
+#### Scenario: Handle complex table structures
+- **WHEN** processing a table with mixed merged and regular cells (e.g., edit3.pdf with 83 cells including 121 merges)
+- **THEN** the system SHALL NOT split merged cells into individual cells
+- **AND** the output cell count SHALL match the actual visual cell count
+- **AND** the rendered PDF SHALL display correct merged cell boundaries
+
+### Requirement: Visual Element Path Preservation
+The system SHALL preserve image paths for all visual element types during OCR conversion.
+
+#### Scenario: Preserve CHART element paths
+- **WHEN** converting PP-StructureV3 output containing CHART elements
+- **THEN** the system SHALL treat CHART as a visual element type
+- **AND** extract saved_path from the element data
+- **AND** include saved_path in the UnifiedDocument content field
+
+#### Scenario: Support all visual element types
+- **WHEN** processing visual elements of types IMAGE, FIGURE, CHART, DIAGRAM, LOGO, or STAMP
+- **THEN** the system SHALL extract saved_path or img_path for each element
+- **AND** preserve path, width, height, and format in content dictionary
+- **AND** enable downstream PDF generation to embed these images
+
+#### Scenario: Fallback path resolution
+- **WHEN** a visual element has multiple path fields (saved_path, img_path)
+- **THEN** the system SHALL prefer saved_path over img_path
+- **AND** fallback to img_path if saved_path is missing
+- **AND** log warning if both paths are missing
+
+### Requirement: Cell Box Coordinate Validation
+The system SHALL validate cell box coordinates from PP-StructureV3 and handle out-of-bounds cases.
+
+#### Scenario: Detect out-of-bounds coordinates
+- **WHEN** processing cell_boxes from PP-StructureV3
+- **THEN** the system SHALL validate each coordinate against page boundaries (0, 0, page_width, page_height)
+- **AND** log tables with coordinates exceeding page bounds
+- **AND** mark affected cells for fallback processing
+
+#### Scenario: Apply CV line detection fallback
+- **WHEN** cell_boxes coordinates are invalid (out of bounds)
+- **THEN** the system SHALL apply OpenCV line detection as fallback
+- **AND** reconstruct table structure from detected lines
+- **AND** include fallback_used flag in table metadata
+
+#### Scenario: Coordinate normalization
+- **WHEN** coordinates are within page bounds but slightly outside table bbox
+- **THEN** the system SHALL clamp coordinates to table boundaries
+- **AND** preserve relative cell positions
+- **AND** ensure no cells overlap after normalization
+
+### Requirement: Decoration Image Filtering
+The system SHALL filter out minimal decoration images that do not contribute meaningful content.
+
+#### Scenario: Filter tiny images by area
+- **WHEN** extracting images from a document
+- **THEN** the system SHALL calculate image area (width x height)
+- **AND** filter out images with area < 200 square pixels
+- **AND** log filtered image count for debugging
+
+#### Scenario: Configurable filtering threshold
+- **WHEN** processing documents with intentionally small images
+- **THEN** the system SHALL support configuration of minimum image area threshold
+- **AND** default to 200 square pixels if not specified
+- **AND** allow threshold = 0 to disable filtering
+
+### Requirement: Covering Image Removal
+The system SHALL remove covering/redaction images from the final output.
+
+#### Scenario: Detect covering rectangles
+- **WHEN** preprocessing a PDF page
+- **THEN** the system SHALL detect black/white rectangles covering text regions
+- **AND** identify covering images by high IoU (> 0.8) with underlying content
+- **AND** mark covering images for exclusion
+
+#### Scenario: Exclude covering images from rendering
+- **WHEN** generating output PDF
+- **THEN** the system SHALL exclude images marked as covering
+- **AND** preserve the text content that was covered
+- **AND** include covering_images_removed count in metadata
+
+#### Scenario: Handle both black and white covering
+- **WHEN** detecting covering rectangles
+- **THEN** the system SHALL detect both black fill (redaction style)
+- **AND** white fill (whiteout style)
+- **AND** low-contrast rectangles intended to hide content
+
+## MODIFIED Requirements
+
+### Requirement: Enhanced OCR with Full PP-StructureV3
+The system SHALL utilize the full capabilities of PP-StructureV3, extracting all 23 element types from parsing_res_list, with proper handling of visual elements and table coordinates.
+
+#### Scenario: Extract comprehensive document structure
+- **WHEN** processing through OCR track
+- **THEN** the system SHALL use page_result.json['parsing_res_list']
+- **AND** extract all element types including headers, lists, tables, figures
+- **AND** preserve layout_bbox coordinates for each element
+
+#### Scenario: Maintain reading order
+- **WHEN** extracting elements from PP-StructureV3
+- **THEN** the system SHALL preserve the reading order from parsing_res_list
+- **AND** assign sequential indices to elements
+- **AND** support reordering for complex layouts
+
+#### Scenario: Extract table structure
+- **WHEN** PP-StructureV3 identifies a table
+- **THEN** the system SHALL extract cell content and boundaries
+- **AND** validate cell_boxes coordinates against page boundaries
+- **AND** apply fallback detection for invalid coordinates
+- **AND** preserve table HTML for structure
+- **AND** extract plain text for translation
+
+#### Scenario: Extract visual elements with paths
+- **WHEN** PP-StructureV3 identifies visual elements (IMAGE, FIGURE, CHART, DIAGRAM)
+- **THEN** the system SHALL preserve saved_path for each element
+- **AND** include image dimensions and format
+- **AND** enable image embedding in output PDF
+
+### Requirement: Generate UnifiedDocument from direct extraction
+The system SHALL convert PyMuPDF results to UnifiedDocument with correct table cell merging.
+
+#### Scenario: Extract tables with cell merging
+- **WHEN** direct extraction encounters a table
+- **THEN** the system SHALL use PyMuPDF find_tables() API
+- **AND** extract cell content with correct rowspan/colspan
+- **AND** preserve merged cell boundaries
+- **AND** skip placeholder cells covered by merges
+
+#### Scenario: Filter decoration images
+- **WHEN** extracting images from PDF
+- **THEN** the system SHALL filter images smaller than minimum area threshold
+- **AND** exclude covering/redaction images
+- **AND** preserve meaningful content images
+
+#### Scenario: Preserve text styling with image handling
+- **WHEN** direct extraction completes
+- **THEN** the system SHALL convert PyMuPDF results to UnifiedDocument
+- **AND** preserve text styling, fonts, and exact positioning
+- **AND** extract tables with cell boundaries, content, and merge info
+- **AND** include only meaningful images in output

openspec/changes/refactor-dual-track-architecture/tasks.md

@@ -0,0 +1,108 @@
+# Tasks: Refactor Dual-Track Architecture
+
+## Phase 1: 修復已知 Bug (已完成)
+
+### 1.1 Direct Track 表格修復 (已完成 ✓)
+- [x] 1.1.1 修改 `_process_native_table()` 方法使用 `table.cells` 處理合併單元格
+- [x] 1.1.2 使用 PyMuPDF `page.find_tables()` API (已在使用中)
+- [x] 1.1.3 解析 `table.cells` 並正確計算 `row_span`/`col_span`
+- [x] 1.1.4 處理被合併的單元格（跳過 `None` 值，建立 covered grid）
+- [x] 1.1.5 驗證 edit3.pdf 返回 83 個正確的 cells ✓
+
+### 1.2 OCR Track 圖片路徑修復 (已完成 ✓)
+- [x] 1.2.1 修改 `ocr_to_unified_converter.py` 第 604-613 行
+- [x] 1.2.2 擴展視覺元素類型判斷：`IMAGE, FIGURE, CHART, DIAGRAM, LOGO, STAMP`
+- [x] 1.2.3 優先使用 `saved_path`，fallback 到 `img_path`
+- [x] 1.2.4 確保 content dict 包含 `saved_path`, `path`, `width`, `height`, `format`
+- [x] 1.2.5 程式碼已修正 (需 OCR Track 完整測試驗證)
+- [x] 1.2.6 程式碼已修正 (需 OCR Track 完整測試驗證)
+
+### 1.3 Cell boxes 座標驗證 (已完成 ✓)
+- [x] 1.3.1 在 `ocr_to_unified_converter.py` 添加 `validate_cell_boxes()` 函數
+- [x] 1.3.2 檢查 cell_boxes 是否超出頁面邊界 (0, 0, page_width, page_height)
+- [x] 1.3.3 超出範圍時使用 clamped coordinates，標記 needs_fallback
+- [x] 1.3.4 添加日誌記錄異常座標
+- [x] 1.3.5 單元測試驗證座標驗證邏輯正確 ✓
+
+### 1.4 過濾極小裝飾圖片 (已完成 ✓)
+- [x] 1.4.1 在 `direct_extraction_engine.py` 圖片提取邏輯添加面積檢查
+- [x] 1.4.2 過濾 `image_area < min_image_area` (默認 200 px²) 的圖片
+- [x] 1.4.3 添加 `min_image_area` 配置項允許調整閾值
+- [x] 1.4.4 驗證 edit3.pdf 偵測到 3 個極小裝飾圖片 ✓
+
+### 1.5 移除覆蓋圖像 (已完成 ✓)
+- [x] 1.5.1 傳遞 `covering_images` 到 `_extract_images()` 方法
+- [x] 1.5.2 使用 IoU 閾值 (0.8) 和 xref 比對判斷覆蓋圖像
+- [x] 1.5.3 從最終輸出中排除覆蓋圖像
+- [x] 1.5.4 添加 `_calculate_iou()` 輔助方法
+- [x] 1.5.5 驗證 edit3.pdf 偵測到 6 個黑框覆蓋圖像 ✓
+
+## Phase 2: 服務層重構
+
+### 2.1 提取 ProcessingOrchestrator
+- [ ] 2.1.1 建立 `backend/app/services/processing_orchestrator.py`
+- [ ] 2.1.2 從 OCRService 提取流程編排邏輯
+- [ ] 2.1.3 定義 `ProcessingPipeline` 介面
+- [ ] 2.1.4 實現 DirectPipeline 和 OCRPipeline
+- [ ] 2.1.5 更新 OCRService 使用 ProcessingOrchestrator
+- [ ] 2.1.6 確保現有功能不受影響
+
+### 2.2 提取 TableRenderer
+- [ ] 2.2.1 建立 `backend/app/services/pdf_table_renderer.py`
+- [ ] 2.2.2 從 PDFGeneratorService 提取 HTMLTableParser
+- [ ] 2.2.3 提取表格渲染邏輯到獨立類
+- [ ] 2.2.4 支援合併單元格渲染
+- [ ] 2.2.5 更新 PDFGeneratorService 使用 TableRenderer
+
+### 2.3 提取 FontManager
+- [ ] 2.3.1 建立 `backend/app/services/pdf_font_manager.py`
+- [ ] 2.3.2 提取字體載入和快取邏輯
+- [ ] 2.3.3 提取 CJK 字體支援邏輯
+- [ ] 2.3.4 實現字體 fallback 機制
+- [ ] 2.3.5 更新 PDFGeneratorService 使用 FontManager
+
+## Phase 3: 記憶體管理簡化
+
+### 3.1 統一記憶體策略引擎
+- [ ] 3.1.1 建立 `backend/app/services/memory_policy_engine.py`
+- [ ] 3.1.2 定義統一的記憶體策略介面
+- [ ] 3.1.3 合併 MemoryManager 和 MemoryGuard 邏輯
+- [ ] 3.1.4 整合 Semaphore 管理
+- [ ] 3.1.5 簡化配置到 3-4 個核心項目
+
+### 3.2 更新服務使用新記憶體引擎
+- [ ] 3.2.1 更新 OCRService 使用 MemoryPolicyEngine
+- [ ] 3.2.2 更新 ServicePool 使用 MemoryPolicyEngine
+- [ ] 3.2.3 移除舊的 MemoryGuard 引用
+- [ ] 3.2.4 驗證 GPU 記憶體監控正常運作
+
+## Phase 4: 前端狀態管理改進
+
+### 4.1 新增 TaskStore
+- [ ] 4.1.1 建立 `frontend/src/store/taskStore.ts`
+- [ ] 4.1.2 定義任務狀態結構（currentTask, tasks, processingStatus）
+- [ ] 4.1.3 實現 CRUD 操作和狀態轉換
+- [ ] 4.1.4 添加 localStorage 持久化
+- [ ] 4.1.5 更新 ProcessingPage 使用 TaskStore
+- [ ] 4.1.6 更新 TaskDetailPage 使用 TaskStore
+
+### 4.2 合併類型定義
+- [ ] 4.2.1 審查 `api.ts` 和 `apiV2.ts` 的差異
+- [ ] 4.2.2 合併類型定義到 `apiV2.ts`
+- [ ] 4.2.3 移除 `api.ts` 中的重複定義
+- [ ] 4.2.4 更新所有 import 路徑
+- [ ] 4.2.5 驗證 TypeScript 編譯無錯誤
+
+## Phase 5: 測試與驗證
+
+### 5.1 回歸測試
+- [ ] 5.1.1 使用 edit.pdf 測試 Direct Track（確保無回歸）
+- [ ] 5.1.2 使用 edit3.pdf 測試 Direct Track 表格合併
+- [ ] 5.1.3 使用 edit.pdf 測試 OCR Track 圖片放回
+- [ ] 5.1.4 使用 edit3.pdf 測試 OCR Track 圖片放回
+- [ ] 5.1.5 驗證所有 cell_boxes 座標正確
+
+### 5.2 效能測試
+- [ ] 5.2.1 測量重構後的處理時間
+- [ ] 5.2.2 驗證記憶體使用無明顯增加
+- [ ] 5.2.3 驗證 GPU 使用率正常