paddle ocr 위치 기반 텍스트 정렬

utils/text_extractor.py

@@ -149,8 +149,8 @@ def extract_text_paddle_ocr(images):
         use_doc_orientation_classify=False, use_doc_unwarping=False, lang="korean"
     )
 
-    full_response = []
     coord_response = []
+    all_text_boxes = []  # (y_center, x_center, text, box) 저장용
 
     for page_idx, img in enumerate(images):
         print(f"[PaddleOCR] 페이지 {page_idx + 1} OCR로 텍스트 추출 중...")
@@ -183,13 +183,50 @@ def extract_text_paddle_ocr(images):
             texts = res_dic.get("rec_texts", [])
             boxes = res_dic.get("rec_boxes", [])
 
-            full_response.extend(texts)
+            for text, box in zip(texts, boxes):
+                if isinstance(box, np.ndarray):
+                    box = box.tolist()
+                # ✅ box 정규화
+                if all(isinstance(p, (int, float)) for p in box):
+                    if len(box) % 2 == 0:
+                        box = [[box[i], box[i + 1]] for i in range(0, len(box), 2)]
+                    else:
+                        print(f"[PaddleOCR] 잘못된 box 형식: {box}")
+                        continue
 
-            # ndarray → list 변환
-            clean_boxes = [
-                box.tolist() if isinstance(box, np.ndarray) else box for box in boxes
-            ]
-            coord_response.extend(clean_boxes)
+                coord_response.append(box)
+
+                # 중심 좌표 계산 (y → 줄 순서, x → 단어 순서)
+                x_coords = [p[0] for p in box]
+                y_coords = [p[1] for p in box]
+                x_center = sum(x_coords) / len(x_coords)
+                y_center = sum(y_coords) / len(y_coords)
+
+                all_text_boxes.append((y_center, x_center, text))
+
+    # ✅ 위치 기반 정렬
+    all_text_boxes.sort(key=lambda x: (x[0], x[1]))  # y 먼저, 그 다음 x 정렬
+
+    # ✅ 줄 단위 그룹핑
+    lines = []
+    current_line = []
+    prev_y = None
+    line_threshold = 15  # 줄 묶음 y 오차 허용값
+
+    for y, x, text in all_text_boxes:
+        if prev_y is None or abs(y - prev_y) < line_threshold:
+            current_line.append((x, text))
+        else:
+            current_line.sort(key=lambda xx: xx[0])
+            lines.append(" ".join(t for _, t in current_line))
+            current_line = [(x, text)]
+        prev_y = y
+
+    if current_line:
+        current_line.sort(key=lambda xx: xx[0])
+        lines.append(" ".join(t for _, t in current_line))
+
+    parsed_text = "\n".join(lines)
 
     print("[PaddleOCR] 전체 페이지 텍스트 및 좌표 추출 완료")
-    return " ".join(full_response), coord_response
+    return parsed_text, coord_response