Commit baba06ae authored by dmMaze's avatar dmMaze
Browse files

fix coordinates overflow of textlines, close #552

parent ecf39bd7

modules/textdetector/ctd/inference.py

+1 −1
Original line number Diff line number Diff line
@@ -343,7 +343,7 @@ class TextDetector: 
        if lines.size == 0:

            lines = []

        else:

            lines = lines.astype(np.int32)

            lines = lines.astype(np.int64)

        blk_list = group_output(blks, lines, im_w, im_h, mask)

        mask_refined = refine_mask(img, mask, blk_list, refine_mode=refine_mode)

        if keep_undetected_mask:

modules/textdetector/db_utils.py

+2 −2
Original line number Diff line number Diff line
@@ -139,7 +139,7 @@ class SegDetectorRepresenter(): 
        height, width = bitmap.shape

        contours, _ = cv2.findContours((bitmap * 255).astype(np.uint8), cv2.RETR_LIST, cv2.CHAIN_APPROX_SIMPLE)

        num_contours = min(len(contours), self.max_candidates)

        boxes = np.zeros((num_contours, 4, 2), dtype=np.int16)

        boxes = np.zeros((num_contours, 4, 2), dtype=np.int64)

        scores = np.zeros((num_contours,), dtype=np.float32)



        for index in range(num_contours):
@@ -165,7 +165,7 @@ class SegDetectorRepresenter(): 


            box[:, 0] = np.clip(np.round(box[:, 0] / width * dest_width), 0, dest_width)

            box[:, 1] = np.clip(np.round(box[:, 1] / height * dest_height), 0, dest_height)

            boxes[index, :, :] = box.astype(np.int16)

            boxes[index, :, :] = box.astype(np.int64)

            scores[index] = score

        return boxes, scores

utils/textblock.py

+15 −3
Original line number Diff line number Diff line
@@ -236,9 +236,21 @@ class TextBlock: 
        return blk_dict



    def get_transformed_region(self, img: np.ndarray, idx: int, textheight: int, maxwidth: int = None) -> np.ndarray :

        

        line = np.round(np.array(self.lines[idx])).astype(np.int64)

        x1, y1, x2, y2 = line[:, 0].min(), line[:, 1].min(), line[:, 0].max(), line[:, 1].max()

        im_h, im_w = img.shape[:2]

        x1 = np.clip(x1, 0, im_w)

        y1 = np.clip(y1, 0, im_h)

        x2 = np.clip(x2, 0, im_w)

        y2 = np.clip(y2, 0, im_h)

        img_croped = img[y1: y2, x1: x2]

        

        direction = 'v' if self.src_is_vertical else 'h'



        src_pts = np.array(self.lines[idx], dtype=np.float64)

        src_pts = line.copy()

        src_pts[:, 0] -= x1

        src_pts[:, 1] -= y1

        middle_pnt = (src_pts[[1, 2, 3, 0]] + src_pts) / 2

        vec_v = middle_pnt[2] - middle_pnt[0]   # vertical vectors of textlines

        vec_h = middle_pnt[1] - middle_pnt[3]   # horizontal vectors of textlines
@@ -264,7 +276,7 @@ class TextBlock: 
            if M is None:

                print('invalid textpolygon to target img')

                return np.zeros((textheight, textheight, 3), dtype=np.uint8)

            region = cv2.warpPerspective(img, M, (w, h))

            region = cv2.warpPerspective(img_croped, M, (w, h))

        elif direction == 'v' :

            w = int(textheight)

            h = int(round(textheight * ratio))
@@ -273,7 +285,7 @@ class TextBlock: 
            if M is None:

                print('invalid textpolygon to target img')

                return np.zeros((textheight, textheight, 3), dtype=np.uint8)

            region = cv2.warpPerspective(img, M, (w, h))

            region = cv2.warpPerspective(img_croped, M, (w, h))

            region = cv2.rotate(region, cv2.ROTATE_90_COUNTERCLOCKWISE)



        if maxwidth is not None: