feat(rapidocr_onnxruntime): 优化单字符坐标及返回单字符置信度 (#293)

* 1、utils.py：优化单字符坐标计算。根据CTC解码特点，增加每个字符长度列表作为判断条件，一般非中文字符2-3个解码宽度，中文3-4个解码宽度，当某个字符与上一字符距离大于4个解码宽度，表示中间有空白，此时进行字段拆分，这样可以减少文字间空白对坐标计算的影响；同时返回每个字符值信度，以方便应用使用

2、cal_rec_boxes/main.py：
 配合返回单字符值信度列表的修改
 修改计算单字符宽度的分母为总字符数减1，与分母一致，这样计算出的值更加准确

 3.main.py：
 修复主程序没有传递命令行参数的问题

* 优化文本行检测优化，当limit_type为max，根据图片大小动态设置limit_side_len，以减少漏检

---------

Co-authored-by: jesse01 <jesse01.163.com>

python/rapidocr_onnxruntime/cal_rec_boxes/main.py

@@ -28,14 +28,14 @@ class CalRecBoxes:
             rec_txt, rec_conf, rec_word_info = rec_res[0], rec_res[1], rec_res[2]
             h, w = img.shape[:2]
             img_box = np.array([[0, 0], [w, 0], [w, h], [0, h]])
-            word_box_content_list, word_box_list = self.cal_ocr_word_box(
+            word_box_content_list, word_box_list, conf_list = self.cal_ocr_word_box(
                 rec_txt, img_box, rec_word_info
             )
             word_box_list = self.adjust_box_overlap(copy.deepcopy(word_box_list))
             word_box_list = self.reverse_rotate_crop_image(
                 copy.deepcopy(box), word_box_list, direction
             )
-            res.append([rec_txt, rec_conf, word_box_list, word_box_content_list])
+            res.append([rec_txt, rec_conf, word_box_list, word_box_content_list, conf_list])
         return res
 
     @staticmethod
@@ -60,13 +60,13 @@ class CalRecBoxes:
     @staticmethod
     def cal_ocr_word_box(
         rec_txt: str, box: np.ndarray, rec_word_info: List[Tuple[str, List[int]]]
-    ) -> Tuple[List[str], List[List[int]]]:
+    ) -> Tuple[List[str], List[List[int]], List[float]]:
         """Calculate the detection frame for each word based on the results of recognition and detection of ocr
         汉字坐标是单字的
         英语坐标是单词级别的
         """
 
-        col_num, word_list, word_col_list, state_list = rec_word_info
+        col_num, word_list, word_col_list, state_list, conf_list = rec_word_info
         box = box.tolist()
         bbox_x_start = box[0][0]
         bbox_x_end = box[1][0]
@@ -84,7 +84,7 @@ class CalRecBoxes:
         def cal_char_width(width_list, word_col_):
             if len(word_col_) == 1:
                 return
-            char_total_length = (word_col_[-1] - word_col_[0] + 1) * cell_width
+            char_total_length = (word_col_[-1] - word_col_[0]) * cell_width
             char_width = char_total_length / (len(word_col_) - 1)
             width_list.append(char_width)
 
@@ -124,7 +124,7 @@ class CalRecBoxes:
         cal_box(cn_col_list, cn_width_list, word_box_list)
         cal_box(en_col_list, en_width_list, word_box_list)
         sorted_word_box_list = sorted(word_box_list, key=lambda box: box[0][0])
-        return word_box_content_list, sorted_word_box_list
+        return word_box_content_list, sorted_word_box_list, conf_list
 
     @staticmethod
     def adjust_box_overlap(

python/rapidocr_onnxruntime/ch_ppocr_det/text_detect.py

@@ -26,9 +26,9 @@ from .utils import DBPostProcess, DetPreProcess
 
 class TextDetector:
     def __init__(self, config: Dict[str, Any]):
-        limit_side_len = config.get("limit_side_len", 736)
-        limit_type = config.get("limit_type", "min")
-        self.preprocess_op = DetPreProcess(limit_side_len, limit_type)
+        self.limit_type = config.get("limit_type", "min")
+        self.limit_side_len = config.get("limit_side_len", 736)
+        self.preprocess_op = None
 
         post_process = {
             "thresh": config.get("thresh", 0.3),
@@ -49,6 +49,7 @@ class TextDetector:
             raise ValueError("img is None")
 
         ori_img_shape = img.shape[0], img.shape[1]
+        self.preprocess_op = self.get_preprocess(max(img.shape[0], img.shape[1]))
         prepro_img = self.preprocess_op(img)
         if prepro_img is None:
             return None, 0
@@ -59,6 +60,17 @@ class TextDetector:
         elapse = time.perf_counter() - start_time
         return dt_boxes, elapse
 
+    def get_preprocess(self, max_wh):
+        if self.limit_type == 'min':
+            limit_side_len = self.limit_side_len
+        elif max_wh < 960:
+            limit_side_len = 960
+        elif max_wh < 1500:
+            limit_side_len = 1500
+        else:
+            limit_side_len = 2000
+        return DetPreProcess(limit_side_len, self.limit_type)
+
     def filter_tag_det_res(
         self, dt_boxes: np.ndarray, image_shape: Tuple[int, int]
     ) -> np.ndarray:

python/rapidocr_onnxruntime/ch_ppocr_rec/utils.py

@@ -92,7 +92,7 @@ class CTCLabelDecode:
                 selection &= text_index[batch_idx] != ignored_token
 
             if text_prob is not None:
-                conf_list = text_prob[batch_idx][selection]
+                conf_list = np.array(text_prob[batch_idx][selection]).tolist()
             else:
                 conf_list = [1] * len(selection)
 
@@ -116,6 +116,7 @@ class CTCLabelDecode:
                             word_list,
                             word_col_list,
                             state_list,
+                            conf_list
                         ],
                     )
                 )
@@ -147,7 +148,11 @@ class CTCLabelDecode:
         word_list = []
         word_col_list = []
         state_list = []
-        valid_col = np.where(selection == True)[0]
+        valid_col = np.where(selection)[0]
+        col_width = np.zeros(valid_col.shape)
+        if len(valid_col) > 0:
+            col_width[1:] = valid_col[1:] - valid_col[:-1]
+            col_width[0] = min(3 if "\u4e00" <= text[0] <= "\u9fff" else 2, int(valid_col[0]))
 
         for c_i, char in enumerate(text):
             if "\u4e00" <= char <= "\u9fff":
@@ -155,10 +160,10 @@ class CTCLabelDecode:
             else:
                 c_state = "en&num"
 
-            if state == None:
+            if state is None:
                 state = c_state
 
-            if state != c_state:
+            if state != c_state or col_width[c_i] > 4:
                 if len(word_content) != 0:
                     word_list.append(word_content)
                     word_col_list.append(word_col_content)

python/rapidocr_onnxruntime/main.py

@@ -334,6 +334,7 @@ def main():
         use_det=use_det,
         use_cls=use_cls,
         use_rec=use_rec,
+        **vars(args)
     )
     logger.info(result)