update rapidocr version

Author: hzkitty

pyproject.toml

@@ -37,7 +37,7 @@ dependencies = [
     "ftfy>=6.3.1,<7",
     "shapely>=2.0.7,<3",
     "tokenizers>=0.13.2",
-    "rapidocr>=3.1.0,<=3.4.2",
+    "rapidocr>=3.4.0,<=3.4.2",
 ]
 
 [project.optional-dependencies]

rapid_doc/model/ocr/rapid_ocr.py

@@ -1,4 +1,4 @@
-from rapidocr.ch_ppocr_det import TextDetOutput
+from typing import List, Dict, Any
 
 from rapid_doc.model.ocr.ocr_patch import apply_ocr_patch
 
@@ -157,13 +157,62 @@ def ocr(self,
                         op_record = {'padding_1': {'left': 0, 'top': 0}, 'preprocess': {'ratio_h': 1.0, 'ratio_w': 1.0}}
                         raw_h, raw_w = ori_img.shape[:2]
                         dt_boxes_np = [np.array(box, dtype=np.float32) for box in dt_boxes]
-                        word_results = self.ocr_engine.calc_word_boxes(img, TextDetOutput(boxes=dt_boxes_np), rec_result, op_record, raw_h, raw_w)
+                        word_results = self.calc_word_boxes(img, dt_boxes_np, rec_result, op_record, raw_h, raw_w)
                         rec_res = list(zip(rec_result.txts, rec_result.scores, word_results))
                     else:
                         rec_res = list(zip(rec_result.txts, rec_result.scores))
                     ocr_res.append(rec_res)
                 return ocr_res
 
+    def calc_word_boxes(
+        self,
+        img: List[np.ndarray],
+        dt_boxes: np.ndarray,
+        rec_res: TextRecOutput,
+        op_record: Dict[str, Any],
+        raw_h: int,
+        raw_w: int,
+    ) -> Any:
+        rec_res = self.ocr_engine.cal_rec_boxes(
+            img, dt_boxes, rec_res, self.ocr_engine.return_single_char_box
+        )
+
+        origin_words = []
+        for word_line in rec_res.word_results:
+            origin_words_item = []
+            for txt, score, bbox in word_line:
+                if bbox is None:
+                    continue
+
+                origin_words_points = self.map_boxes_to_original(
+                    np.array([bbox]).astype(np.float64), op_record, raw_h, raw_w
+                )
+                origin_words_points = origin_words_points.astype(np.int32).tolist()[0]
+                origin_words_item.append((txt, score, origin_words_points))
+
+            if origin_words_item:
+                origin_words.append(tuple(origin_words_item))
+        return tuple(origin_words)
+
+    def map_boxes_to_original(
+            self, dt_boxes: np.ndarray, op_record: Dict[str, Any], ori_h: int, ori_w: int
+    ) -> np.ndarray:
+        for op in reversed(list(op_record.keys())):
+            v = op_record[op]
+            if "padding" in op:
+                top, left = v.get("top"), v.get("left")
+                dt_boxes[:, :, 0] -= left
+                dt_boxes[:, :, 1] -= top
+            elif "preprocess" in op:
+                ratio_h = v.get("ratio_h")
+                ratio_w = v.get("ratio_w")
+                dt_boxes[:, :, 0] *= ratio_w
+                dt_boxes[:, :, 1] *= ratio_h
+
+        dt_boxes = np.where(dt_boxes < 0, 0, dt_boxes)
+        dt_boxes[..., 0] = np.where(dt_boxes[..., 0] > ori_w, ori_w, dt_boxes[..., 0])
+        dt_boxes[..., 1] = np.where(dt_boxes[..., 1] > ori_h, ori_h, dt_boxes[..., 1])
+        return dt_boxes
 
     def __call__(self, img, mfd_res=None):