Is it possible to get the angle of rotation of the text area via PaddleOCR?

[Aflexg]

PaddleOCR can correctly recognize 90, 180 and even 270 degree rotated text in a mode use_angle_cls=True, but it doesnt provide any information about the angle of rotation in a result.As a result, area with a text, cropped by provided coordinates usually has incorrect orientation.Can i get the information about the angle of rotation of the text area somehow to get correctly orientated area witht text ?

[UserWangZz]

Specifically you need to look at the original code to determine the output of the classification model.

[Aflexg]

To be honest i cant understand how does this code work for now.I am compeltly sure that it must rotate part of the image with a text if it is not horizontal but i cant find any image rotation operations in the code.

[UserWangZz]

You can see this part code in paddleocr.py, if you use paddleocr api. About 755 line, use a classifier model to det the det model's result. and then you can find second part code in tools/infer/predict_cls.py. Here use cls model to rectify det box.

PaddleOCR/paddleocr.py

Lines 725 to 762 in 7a3c580

	if det and rec:
	ocr_res = []
	for idx, img in enumerate(imgs):
	img = preprocess_image(img)
	dt_boxes, rec_res, _ = self.__call__(img, cls, slice)
	if not dt_boxes and not rec_res:
	ocr_res.append(None)
	continue
	tmp_res = [[box.tolist(), res] for box, res in zip(dt_boxes, rec_res)]
	ocr_res.append(tmp_res)
	return ocr_res
	elif det and not rec:
	ocr_res = []
	for idx, img in enumerate(imgs):
	img = preprocess_image(img)
	dt_boxes, elapse = self.text_detector(img)
	if dt_boxes.size == 0:
	ocr_res.append(None)
	continue
	tmp_res = [box.tolist() for box in dt_boxes]
	ocr_res.append(tmp_res)
	return ocr_res
	else:
	ocr_res = []
	cls_res = []
	for idx, img in enumerate(imgs):
	if not isinstance(img, list):
	img = preprocess_image(img)
	img = [img]
	if self.use_angle_cls and cls:
	img, cls_res_tmp, elapse = self.text_classifier(img)
	if not rec:
	cls_res.append(cls_res_tmp)
	rec_res, elapse = self.text_recognizer(img)
	ocr_res.append(rec_res)
	if not rec:
	return cls_res
	return ocr_res

PaddleOCR/tools/infer/predict_cls.py

Lines 119 to 126 in 7a3c580

	for rno in range(len(cls_result)):
	label, score = cls_result[rno]
	cls_res[indices[beg_img_no + rno]] = [label, score]
	if "180" in label and score > self.cls_thresh:
	img_list[indices[beg_img_no + rno]] = cv2.rotate(
	img_list[indices[beg_img_no + rno]], 1
	)
	return img_list, cls_res, elapse

[Aflexg]

Thank you very much for the provided information. With your help, I was able to get all the necessary information about the rotation angles of the text areas in the image.

[Aagam2]

Hello @Aflexg how did you get information about angles besides 0 and 180. In the model configuration when I specify label_list=['0', '90', '180', '270']
The model's results are not correct.

[Aflexg]

Hello @Aagam2. I’m not sure that I get information about angles other than 0 and 180. All I did was make minor modifications to the paddleocr.py script so that it outputs data about rotation angles using the information mentioned above.

[Aagam2]

Yes I was able to get information about the angles after making modification but it is only working for 0 and 180 degree rotated image.
Thank you for the reply.