improve the OCR pipeline processing and outputs formatting

src/ocr_pipeline/detectors/figures.py

@@ -33,7 +33,10 @@ class FigureDetector:
         try:
             import layoutparser as lp
 
-            self._predictor = lp.Detectron2LayoutModel(
+            model_class = getattr(lp, "Detectron2LayoutModel", None) or getattr(lp, "AutoLayoutModel", None)
+            if model_class is None:
+                raise RuntimeError("layoutparser has no compatible layout model backend; install Detectron2")
+            self._predictor = model_class(
                 config_path=settings.detectors.figures.model,
                 label_map={0: "Text", 1: "Title", 2: "List", 3: "Table", 4: "Figure"},
                 extra_config=[
@@ -67,7 +70,10 @@ class FigureDetector:
                         block.block.x_2,
                         block.block.y_2,
                     )
-                    crop = image[int(y1) : int(y2), int(x1) : int(x2)]
+                    height, width = image.shape[:2]
+                    x1_i, x2_i = max(0, int(x1)), min(width, int(x2))
+                    y1_i, y2_i = max(0, int(y1)), min(height, int(y2))
+                    crop = image[y1_i:y2_i, x1_i:x2_i]
                     figures.append(
                         DetectedFigure(
                             bbox=[float(x1), float(y1), float(x2), float(y2)],
@@ -98,7 +104,10 @@ class CaptionExtractor:
         try:
             import layoutparser as lp
 
-            self._predictor = lp.Detectron2LayoutModel(
+            model_class = getattr(lp, "Detectron2LayoutModel", None) or getattr(lp, "AutoLayoutModel", None)
+            if model_class is None:
+                raise RuntimeError("layoutparser has no compatible layout model backend; install Detectron2")
+            self._predictor = model_class(
                 config_path=settings.detectors.captions.model,
                 label_map={0: "Text", 1: "Title", 2: "List", 3: "Table", 4: "Figure"},
                 extra_config=[
@@ -156,11 +165,14 @@ class CaptionExtractor:
             return figures
 
 
-def detect_figures(image: np.ndarray) -> list[DetectedFigure]:
-    detector = FigureDetector()
-    figures = detector.detect(image)
+_figure_detector = FigureDetector()
+_caption_extractor = CaptionExtractor()
 
-    captioner = CaptionExtractor()
+
+def detect_figures(image: np.ndarray) -> list[DetectedFigure]:
+    figures = _figure_detector.detect(image)
+
+    captioner = _caption_extractor
     figures = captioner.extract_captions(image, figures)
 
     return figures

src/ocr_pipeline/detectors/tables.py

@@ -59,6 +59,7 @@ class TableDetector:
         try:
             import torch
 
+            original_height, original_width = image.shape[:2]
             pil_image = cv2.cvtColor(image, cv2.COLOR_BGR2RGB)
             pil_image = cv2.resize(pil_image, (800, 800))
 
@@ -77,8 +78,12 @@ class TableDetector:
             for score, _label, box in zip(
                 results["scores"], results["labels"], results["boxes"], strict=True
             ):
-                box = [float(x) for x in box]
+                resized_box = [float(x) for x in box]
+                scale_x, scale_y = original_width / 800, original_height / 800
+                box = [resized_box[0] * scale_x, resized_box[1] * scale_y, resized_box[2] * scale_x, resized_box[3] * scale_y]
                 x1, y1, x2, y2 = map(int, box)
+                x1, x2 = max(0, x1), min(original_width, x2)
+                y1, y2 = max(0, y1), min(original_height, y2)
                 crop = image[y1:y2, x1:x2]
                 tables.append(
                     DetectedTable(
@@ -122,6 +127,7 @@ class TableStructureRecognizer:
             self._initialized = True
 
     def recognize(self, table_crop: np.ndarray) -> tuple[list[dict[str, Any]] | None, str | None]:
+        self._init()
         if self._model is None or self._processor is None:
             return None, None
 
@@ -150,8 +156,9 @@ class TableStructureRecognizer:
                     }
                 )
 
-            markdown = self._structure_to_markdown(structure, table_crop.shape)
-            return structure, markdown
+            # Structure models locate rows/cells but do not provide cell text. Do not
+            # emit labels as fabricated table content; OCR must be associated separately.
+            return structure, None
         except Exception as e:
             logger.warning("table_structure_recognition_failed", error=str(e))
             return None, None
@@ -185,17 +192,19 @@ class TableStructureRecognizer:
         return None
 
 
+_table_detector = TableDetector()
+_table_structure_recognizer = TableStructureRecognizer()
+
+
 def detect_tables(image: np.ndarray) -> list[DetectedTable]:
-    detector = TableDetector()
-    tables = detector.detect(image)
+    tables = _table_detector.detect(image)
 
     if not tables:
         return tables
 
-    recognizer = TableStructureRecognizer()
     for table in tables:
         if table.image_crop is not None:
-            structure, markdown = recognizer.recognize(table.image_crop)
+            structure, markdown = _table_structure_recognizer.recognize(table.image_crop)
             table.structure = structure
             table.markdown = markdown