search

breezedeus / Pix2Text

An Open-Source Python3 tool with SMALL models for recognizing layouts, tables, math formulas (LaTeX), and text in images, converting them into Markdown format. A free alternative to Mathpix, empowering seamless conversion of visual content into text-based representations. 80+ languages are supported.
https://p2t.breezedeus.com
MIT License
1.99k stars 191 forks source link

one more problem i saw is that the text output is different if I use raw easyocr and your pix2tex.recognize_text_formula #133

Open KTBsomen opened 4 months ago

KTBsomen commented 4 months ago

bug

here is pix2tex output :

73。 What is the compound interest (i Rs) on a sum Of Rs. 8192 for
$$
1 {\frac{1} {4}} \ \mathrm{y e a r s \; a t \; 1 5 \% \; p e r \; a n n u m, \; i f}
$$
"" coooounoed #~ nonths4p (2)  1740 1640
1634 (4)1735
Exassccotzoeolshier]
13.08.2020 (Shift-J)

here is easyocr output:

[([[107, 71], [173, 71], [173, 111], [107, 111]], '73.', 0.9917022049101489), ([[193, 69], [431, 69], [431, 130], [193, 130]], 'What is the', 0.9887154965614935), ([[427, 93], [797, 93], [797, 149], [427, 149]], 'compound interest', 0.9978926405302552), ([[192, 120], [799, 120], [799, 198], [192, 198]], '(in Rs) on a sum of Rs, 8192 for', 0.5171315037820216), ([[237, 203], [257, 203], [257, 231], [237, 231]], '1', 0.09841237729462193), ([[202, 226], [226, 226], [226, 258], [202, 258]], '1', 0.9996012846778548), ([[234, 260], [264, 260], [264, 294], [234, 294]], '4', 1.0), ([[277, 238], [795, 238], [795, 303], [277, 303]], 'years at 15% per annum; if', 0.7160528684004921), ([[184, 312], [369, 312], [369, 365], [184, 365]], 'interest', 0.9999773248890145), ([[380, 326], [722, 326], [722, 383], [380, 383]], 'is compounded', 0.9875069147908678), ([[737, 341], [775, 341], [775, 381], [737, 381]], '5-', 0.8094714373744847), ([[185, 357], [386, 357], [386, 423], [185, 423]], 'monthly ?', 0.8475277887695424), ([[205, 415], [259, 415], [259, 459], [205, 459]], '(1)', 0.9965351652399872), ([[282, 415], [391, 415], [391, 465], [282, 465]], '1640', 0.9999983310699463), ([[506, 428], [679, 428], [679, 477], [506, 477]], '(2) 1740', 0.7805645267010147), ([[203, 467], [257, 467], [257, 511], [203, 511]], '(3)', 0.9998854752209928), ([[283, 471], [385, 471], [385, 515], [283, 515]], '1634', 0.6752449204900562), ([[502, 480], [677, 480], [677, 530], [502, 530]], '(4) 1735', 0.8574826030860164), ([[401, 523], [786, 523], [786, 584], [401, 584]], 'SSC CGL (CBE) Tier-I', 0.8960124626829515), ([[448, 574], [782, 574], [782, 628], [448, 628]], '13.08.2020 (Shift-I)', 0.7733794959921725), ([[335.0939460213589, 569.1039678918432], [449.9501978735121, 585.3707649405025], [443.9060539786411, 619.8960321081568], [329.0498021264879, 604.6292350594975]], 'Exam,', 0.9631557823077296)]

extracted texts are perfect in this raw easyocr

and I am using easyocr I know as I have changed the prepare_ocr_engine function

def prepare_ocr_engine(languages: Sequence[str], ocr_engine_config):
    ocr_engine_config = deepcopy(ocr_engine_config) if ocr_engine_config else {}
    if 1==2:
        from cnocr import CnOcr

        if 'ch_sim' not in languages and 'cand_alphabet' not in ocr_engine_config:  # only recognize english characters
            ocr_engine_config['cand_alphabet'] = string.printable
        ocr_engine = CnOcr(**ocr_engine_config)
        engine_wrapper = CnOCREngine(languages, ocr_engine)
    else:
        print("using easyocr")
        try:
            from easyocr import Reader
        except:
            raise ImportError('Please install easyocr first: pip install easyocr')
        gpu = False
        if 'context' in ocr_engine_config:
            context = ocr_engine_config.pop('context').lower()
            gpu = 'gpu' in context or 'cuda' in context
        ocr_engine = Reader(lang_list=['en'], gpu=gpu, **ocr_engine_config)
        engine_wrapper = EasyOCREngine(languages, ocr_engine)
    return engine_wrapper

Originally posted by @KTBsomen in https://github.com/breezedeus/Pix2Text/issues/130#issuecomment-2232664131

here is the image

sd