Method parse_ocr

caption_anything/model.py:270–291

(self, image, thres=0.2)

        return ','.join(dense_cap_prompt)

    def parse_ocr(self, image, thres=0.2):
        width, height = get_image_shape(image)
        image = load_image(image, return_type='numpy')
        bounds = self.ocr_reader.readtext(image)
        bounds = [bound for bound in bounds if bound[2] > thres]
        print('Process OCR Text:\n', bounds)

        ocr_prompt = []
        for box, text, conf in bounds:
            p0, p1, p2, p3 = box
            ocr_prompt.append('(\"{}\": X:{:.0f}, Y:{:.0f})'.format(text, (p0[0] + p1[0] + p2[0] + p3[0]) / 4,
                                                                    (p0[1] + p1[1] + p2[1] + p3[1]) / 4))
        ocr_prompt = '\n'.join(ocr_prompt)

        # ocr_prompt = self.text_refiner.llm(f'The image have some scene texts with their locations: {ocr_prompt}. Please group these individual words into one or several phrase based on their relative positions (only give me your answer, do not show explanination)').strip()

        # ocr_prefix1 = f'The image have some scene texts with their locations: {ocr_prompt}. Please group these individual words into one or several phrase based on their relative positions (only give me your answer, do not show explanination)'
        # ocr_prefix2 = f'Please group these individual words into 1-3 phrases, given scene texts with their locations: {ocr_prompt}. You return is one or several strings and infer their locations. (only give me your answer like (“man working”, X: value, Y: value), do not show explanination)'
        # ocr_prefix4 = f'summarize the individual scene text words detected by OCR tools into a fluent sentence based on their positions and distances. You should strictly describe all of the given scene text words. Do not miss any given word. Do not create non-exist words. Do not appear numeric positions. The individual words are given:\n{ocr_prompt}\n'
        # ocr_prefix3 = f'combine the individual scene text words detected by OCR tools into one/several fluent phrases/sentences based on their positions and distances. You should strictly copy or correct all of the given scene text words. Do not miss any given word. Do not create non-exist words. The response is several strings seperate with their location (X, Y), each of which represents a phrase. The individual words are given:\n{ocr_prompt}\n'
        # response = self.text_refiner.llm(ocr_prefix3).strip() if len(ocr_prompt) else ""
        return ocr_prompt

    def inference_cap_everything(self, image, verbose=False):
        image = load_image(image, return_type='pil')