{"id":2928,"date":"2023-07-14T18:30:01","date_gmt":"2023-07-14T18:30:01","guid":{"rendered":"https:\/\/visimatic.cz\/?p=2928"},"modified":"2023-07-15T10:39:03","modified_gmt":"2023-07-15T10:39:03","slug":"ocr-s-neuronovou-siti","status":"publish","type":"post","link":"https:\/\/visimatic.cz\/en\/ocr-s-neuronovou-siti\/","title":{"rendered":"OCR s neuronovou s\u00edt\u00ed: revoluce v \u0159e\u0161en\u00ed spolehliv\u00e9 kontroly textu"},"content":{"rendered":"
\n\t\t\t\t\t\t
\n\t\t\t\t\t\t
\n\t\t\t\t\t
\n\t\t\t
\n\t\t\t\t\t\t
\n\t\t\t\t
\n\t\t\t\t\t

OCR: \u010cten\u00ed textu v pr\u016fmyslov\u00fdch aplikac\u00edch pomoc\u00ed deep learning neuronov\u00fdch s\u00edt\u00ed<\/h2>\t\t\t\t<\/div>\n\t\t\t\t<\/div>\n\t\t\t\t\t<\/div>\n\t\t<\/div>\n\t\t\t\t\t<\/div>\n\t\t<\/section>\n\t\t\t\t
\n\t\t\t\t\t\t
\n\t\t\t\t\t
\n\t\t\t
\n\t\t\t\t\t\t
\n\t\t\t\t
\n\t\t\t\t\t\t\t\t\t

P\u00e1r slov \u00favodem…<\/h2>

I p\u0159es to, \u017ee se v pr\u016fmyslu k identifikaci v\u00fdrobk\u016f st\u00e1le \u010dast\u011bji vyu\u017e\u00edv\u00e1 1D\/2D k\u00f3d\u016f (\u010d\u00e1rov\u00e9 k\u00f3dy, QR, DMC aj.), \u010dlov\u011bkem \u010diteln\u00fd text je v n\u011bkter\u00fdch m\u00edstech st\u00e1le zcela nenahraditeln\u00fd. A\u0165 u\u017e jde o interpretaci k\u00f3dovan\u00fdch dat pro p\u0159\u00edpad ne\u010ditelnosti k\u00f3du samotn\u00e9ho, nebo jako informace pro koncov\u00e9ho spot\u0159ebitele (nap\u0159. datum spot\u0159eby), b\u011b\u017en\u00fd text m\u00e1 a nejsp\u00ed\u0161 je\u0161t\u011b n\u011bjakou dobu bude m\u00edt sv\u00e9 nenahraditeln\u00e9 m\u00edsto. Opravdu spolehliv\u00e9 rozpozn\u00e1n\u00ed a kontrola ti\u0161t\u011bn\u00fdch text\u016f metodou OCR je p\u0159itom v re\u00e1ln\u00fdch podm\u00ednk\u00e1ch v\u00fdrobn\u00edch proces\u016f \u010dasto t\u00e9m\u011b\u0159 nemo\u017en\u00e1 \u00faloha. Tedy alespo\u0148 je\u0161t\u011b do ned\u00e1vna byla. Se st\u00e1le v\u011bt\u0161\u00edm p\u0159esahem a rozvojem um\u011bl\u00e9 inteligence a neuronov\u00fdch s\u00edt\u00ed do oblasti zpracov\u00e1n\u00ed obrazu p\u0159ich\u00e1zej\u00ed d\u0159\u00edve nev\u00eddan\u00e9 mo\u017enosti kontroly, jako tomu je i v p\u0159\u00edpad\u011b, kter\u00fd v\u00e1m dnes p\u0159edstav\u00edme.<\/p>

\u010cten\u00ed a rozpozn\u00ed textu v obraze (tzv. OCR – Optical Character Recognition) m\u016f\u017ee b\u00fdt slo\u017eit\u011bj\u0161\u00ed ne\u017e se na prvn\u00ed pohled zd\u00e1. P\u0159i vyu\u017eit\u00ed b\u011b\u017en\u00fdch algoritm\u016f zpracov\u00e1n\u00ed obrazu vych\u00e1z\u00ed cel\u00fd postup v\u011bt\u0161inou z n\u011bkolika samostatn\u00fdch navazuj\u00edc\u00edch krok\u016f:<\/p>

  1. Lokalizace oblasti textu v obraze nap\u0159. pomoc\u00ed funkc\u00ed, vyhled\u00e1vaj\u00edc\u00edch p\u0159edem definovan\u00fd vzor<\/li>
  2. Segmentace textu – odd\u011blen\u00ed obrazov\u00fdch bod\u016f jednotliv\u00fdch p\u00edsmen a \u010d\u00edslic od pozad\u00ed<\/li>
  3. Optick\u00e9 rozpozn\u00e1n\u00ed znak\u016f (OCR) – p\u0159evod obrazov\u00fdch dat reprezentuj\u00edc\u00ed jednotliv\u00e9 znaky na textovou informaci porovn\u00e1n\u00edm s p\u0159edem p\u0159ipraven\u00fdm modelem fontu nebo na z\u00e1klad\u011b vytvo\u0159en\u00ed vlastn\u00edho<\/li><\/ol>
    \u00a0<\/div>\t\t\t\t\t\t\t\t<\/div>\n\t\t\t\t<\/div>\n\t\t\t\t\t<\/div>\n\t\t<\/div>\n\t\t\t\t\t<\/div>\n\t\t<\/section>\n\t\t\t\t
    \n\t\t\t\t\t\t
    \n\t\t\t\t\t
    \n\t\t\t
    \n\t\t\t\t\t\t
    \n\t\t\t\t
    \n\t\t\t\t\t\t\t\t\t\t\t\t
    \n\t\t\t\t\t\t\t\t\t\t\"Three\t\t\t\t\t\t\t\t\t\t\t
    Jednotliv\u00e9 kroky zpracov\u00e1n\u00ed obrazu od lokalizace p\u0159es segmentaci a\u017e po \u00fasp\u011b\u0161n\u00e9 rozpozn\u00e1n\u00ed textu.<\/figcaption>\n\t\t\t\t\t\t\t\t\t\t<\/figure>\n\t\t\t\t\t\t\t\t\t<\/div>\n\t\t\t\t<\/div>\n\t\t\t\t\t<\/div>\n\t\t<\/div>\n\t\t\t\t\t<\/div>\n\t\t<\/section>\n\t\t\t\t
    \n\t\t\t\t\t\t
    \n\t\t\t\t\t
    \n\t\t\t
    \n\t\t\t\t\t\t
    \n\t\t\t\t
    \n\t\t\t\t\t\t\t\t\t

    Co se na tom m\u016f\u017ee pokazit, kdy\u017e je to tak jednoduch\u00e9?<\/h2>

    V re\u00e1ln\u00e9m prost\u0159ed\u00ed v\u00fdrobn\u00edch linek vstupuj\u00ed do procesu zna\u010den\u00ed a n\u00e1sledn\u00e9 kontroly r\u016fzn\u00e9 vlivy okol\u00ed jako jsou vibrace, pohyb produktu, nep\u0159esn\u00e9 polohov\u00e1n\u00ed p\u0159i zna\u010den\u00ed a n\u00e1sledn\u00e9m po\u0159\u00edzen\u00ed sn\u00edmku, okoln\u00ed osv\u011btlen\u00ed, kvalita zna\u010den\u00ed, nehomogenita podkladov\u00e9ho povrchu a mnoho dal\u0161\u00edch. V\u0161echny tyto vlivy maj\u00ed men\u0161\u00ed nebo v\u011bt\u0161\u00ed dopad na v\u00fdsledek, ale jen z\u0159\u00eddka se navz\u00e1jem jejich p\u016fsoben\u00ed vyru\u0161\u00ed, naopak sp\u00ed\u0161e se jejich vlivy s\u010d\u00edtaj\u00ed. Ne v\u017edy je p\u0159itom mo\u017en\u00e9 tyto parazitn\u00ed vlivy eliminovat nebo p\u0159edv\u00eddat a n\u011bkdy je pot\u0159eba se zkr\u00e1tka sm\u00ed\u0159it s t\u00edm, \u017ee budou do procesu zpracov\u00e1n\u00ed obrazu negativn\u011b p\u0159isp\u00edvat.\u00a0<\/p>

    V prv\u00e9 \u0159ad\u011b m\u016f\u017ee selhat samotn\u00e9 vyhled\u00e1n\u00ed oblasti textu, kter\u00e9 je nezbytn\u00e9, pokud doch\u00e1z\u00ed mezi jednotliv\u00fdmi produkty (a jejich sn\u00edmky) k posuvu textu v r\u00e1mci sn\u00edmk\u016f. Tento krok je v\u011bt\u0161inou nevyhnuteln\u00fd a jen m\u00e1lokdy je text v\u017edy na stejn\u00e9 pozici, p\u0159\u00edpadn\u011b jen s minim\u00e1ln\u00edm posunem proti referen\u010dn\u00ed pozici. Vyhled\u00e1n\u00ed oblasti textu p\u0159itom pracuje na principu porovn\u00e1n\u00ed p\u0159edem nastaven\u00e9 p\u0159edlohy (v\u011bt\u0161inou je pou\u017eit hranov\u00fd model kontur textu), kdy v r\u00e1mci aktu\u00e1ln\u00edho obrazu hled\u00e1 nejv\u011bt\u0161\u00ed m\u00edru shody s touto p\u0159edlohou. Pokud je shoda n\u00edzk\u00e1 nap\u0159. vlivem odlesk\u016f od povrchu nebo zhor\u0161en\u00e9 kvality tisku, nelze spolehliv\u011b ur\u010dit p\u0159esnou pozici oblasti textu a n\u00e1sledn\u00e1 segmentace textu a OCR nem\u00e1 d\u00e1le p\u0159\u00edli\u0161 smysl.<\/p>

    I kdy\u017e se poda\u0159\u00ed text \u00fasp\u011b\u0161n\u011b lokalizovat, nen\u00ed je\u0161t\u011b zdaleka vyhr\u00e1no. M\u016f\u017eeme ale u\u017e p\u0159istoupit k segmentaci textu, kter\u00e1 m\u00e1 za c\u00edl odd\u011blit pop\u0159ed\u00ed (pixely pat\u0159\u00edc\u00ed textu) od pozad\u00ed (pixely podkladu). Jak je tomu u v\u011bt\u0161iny \u00faloh zpracov\u00e1n\u00ed obrazu, podstatn\u00fd je zde hlavn\u011b kontrast, tedy rozd\u00edl v intenzit\u011b pixel\u016f pop\u0159ed\u00ed a pozad\u00ed. V\u011bt\u0161ina segmenta\u010dn\u00edch funkc\u00ed tak\u00e9 pracuje s logikou, \u017ee extrahovan\u00fd text je v\u017edy jedn\u00e9, p\u0159edem ur\u010den\u00e9 polarity, tedy bu\u010f jde o tmav\u00fd text na sv\u011btl\u00e9m pozad\u00ed nebo obr\u00e1cen\u011b. Zna\u010dnou v\u00fdzvou je, pokud se polarita st\u0159\u00edd\u00e1 a to je\u0161t\u011b dop\u0159edu nezn\u00e1m\u00fdm zp\u016fsobem. I kdy\u017e je ale polarita konzistentn\u00ed, m\u016f\u017ee doch\u00e1zet k probl\u00e9mu s prom\u011bnlivou podobou pozad\u00ed, kde jedna \u010d\u00e1st textu je v\u016f\u010di sv\u00e9mu pozad\u00ed kontrastn\u00ed a o kousek d\u00e1l je ve sn\u00edmku odlesk nebo st\u00edn, kter\u00fd kontrast zcela pokaz\u00ed. M\u016f\u017ee tak\u00e9 doch\u00e1zet k probl\u00e9m\u016fm s rozd\u011blen\u00edm souvisl\u00e9ho textu na jednotliv\u00e9 znaky, pokud jsou p\u00edsmena nebo \u010d\u00edslice p\u0159\u00edli\u0161 bl\u00edzko sebe a doch\u00e1z\u00ed na \u00farovni obrazov\u00fdch bod\u016f k jejich spl\u00fdv\u00e1n\u00ed v jeden znak. Z cel\u00e9ho procesu OCR je tak spolehliv\u00e1 segmentace \u010dasto ten nejn\u00e1ro\u010dn\u011bj\u0161\u00ed krok.<\/p>\t\t\t\t\t\t\t\t<\/div>\n\t\t\t\t<\/div>\n\t\t\t\t\t<\/div>\n\t\t<\/div>\n\t\t\t\t\t<\/div>\n\t\t<\/section>\n\t\t\t\t

    \n\t\t\t\t\t\t
    \n\t\t\t\t\t
    \n\t\t\t
    \n\t\t\t\t\t\t
    \n\t\t\t\t
    \n\t\t\t\t\t\t\t\t\t\t\t\t
    \n\t\t\t\t\t\t\t\t\t\t\"\"\t\t\t\t\t\t\t\t\t\t\t
    \u00dasp\u011b\u0161n\u00e1 segmentace textu - jedna polarita v\u0161ech znak\u016f a homogenn\u00ed pozad\u00ed.<\/figcaption>\n\t\t\t\t\t\t\t\t\t\t<\/figure>\n\t\t\t\t\t\t\t\t\t<\/div>\n\t\t\t\t<\/div>\n\t\t\t\t\t<\/div>\n\t\t<\/div>\n\t\t\t\t
    \n\t\t\t
    \n\t\t\t\t\t\t
    \n\t\t\t\t
    \n\t\t\t\t\t\t\t\t\t\t\t\t
    \n\t\t\t\t\t\t\t\t\t\t\"\"\t\t\t\t\t\t\t\t\t\t\t
    Ne\u00fasp\u011b\u0161n\u00e1 segmentace textu - p\u00edsmeno \"T\" a \u010d\u00edslice \"2\" a \"3\" zasahuj\u00ed do prolisu v plastu. P\u00edsmena \"W\", \"F\" a \"L\" m\u011bn\u00ed vlivem odlesku polaritu a jejich kontrast k pozad\u00ed dramaticky kles\u00e1.<\/figcaption>\n\t\t\t\t\t\t\t\t\t\t<\/figure>\n\t\t\t\t\t\t\t\t\t<\/div>\n\t\t\t\t<\/div>\n\t\t\t\t\t<\/div>\n\t\t<\/div>\n\t\t\t\t\t<\/div>\n\t\t<\/section>\n\t\t\t\t
    \n\t\t\t\t\t\t
    \n\t\t\t\t\t
    \n\t\t\t
    \n\t\t\t\t\t\t
    \n\t\t\t\t
    \n\t\t\t\t\t\t\t\t\t

    Spr\u00e1vn\u011b lokalizovan\u00fd a kvalitn\u011b segmentovan\u00fd text, rozd\u011bl\u011bn\u00fd na jednotliv\u00e9 znaky je kone\u010dn\u011b vstupem do t\u0159et\u00ed \u010d\u00e1sti procesu – optick\u00e9ho rozpozn\u00e1n\u00ed znak\u016f. V t\u00e9to f\u00e1zi doch\u00e1z\u00ed k samotn\u00e9mu porovn\u00e1n\u00ed extrahovan\u00e9ho znaku s datab\u00e1z\u00ed dan\u00e9ho fontu a ur\u010den\u00ed nejv\u011bt\u0161\u00ed m\u00edry shody. V principu se tento krok podob\u00e1 prvn\u00ed \u010d\u00e1sti vyhled\u00e1n\u00ed cel\u00e9ho textu podle vzoru, ale v re\u00e1lu jsou zde ur\u010dit\u00e9 odli\u0161nosti v zp\u016fsobu porovn\u00e1n\u00ed. Probl\u00e9m m\u016f\u017ee nastat p\u0159i velk\u00e9 podobnosti znak\u016f v r\u00e1mci zvolen\u00e9ho fontu – nap\u0159. b\u00fdv\u00e1 problematick\u00e9 odli\u0161en\u00ed p\u00edsmena „O“ a \u010d\u00edslice „nula“. Podobn\u00fdch potenci\u00e1ln\u00edch z\u00e1m\u011bn existuje cel\u00e1 \u0159ada, a proto existuj\u00ed specializovan\u00e9 fonty pro OCR (OCR-A, OCR-B), kter\u00e9 byly vyvinuty s c\u00edlem lehko zam\u011bniteln\u00e9 znaky jednozna\u010dn\u011b odli\u0161it. Pokud je to tedy esteticky p\u0159ijateln\u00e9, v\u017edy doporu\u010dujeme pou\u017e\u00edvat n\u011bjak\u00fd typ\u00a0 OCR specializovan\u00e9ho fontu.<\/p>\t\t\t\t\t\t\t\t<\/div>\n\t\t\t\t<\/div>\n\t\t\t\t\t<\/div>\n\t\t<\/div>\n\t\t\t\t\t<\/div>\n\t\t<\/section>\n\t\t\t\t

    \n\t\t\t\t\t\t
    \n\t\t\t\t\t
    \n\t\t\t
    \n\t\t\t\t\t\t
    \n\t\t\t\t
    \n\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\"OCR-A_char_unaccented_capital_letters.svg\"\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t<\/div>\n\t\t\t\t<\/div>\n\t\t\t\t\t<\/div>\n\t\t<\/div>\n\t\t\t\t\t<\/div>\n\t\t<\/section>\n\t\t\t\t
    \n\t\t\t\t\t\t
    \n\t\t\t\t\t
    \n\t\t\t
    \n\t\t\t\t\t\t
    \n\t\t\t\t
    \n\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\"\"\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t<\/div>\n\t\t\t\t<\/div>\n\t\t\t\t\t<\/div>\n\t\t<\/div>\n\t\t\t\t
    \n\t\t\t
    \n\t\t\t\t\t\t
    \n\t\t\t\t
    \n\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\"\"\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t<\/div>\n\t\t\t\t<\/div>\n\t\t\t\t\t<\/div>\n\t\t<\/div>\n\t\t\t\t\t<\/div>\n\t\t<\/section>\n\t\t\t\t
    \n\t\t\t\t\t\t
    \n\t\t\t\t\t
    \n\t\t\t
    \n\t\t\t\t\t\t
    \n\t\t\t\t
    \n\t\t\t\t\t\t\t\t\t

    OCR-A byl poprv\u00e9 vyu\u017eit\u00fd ji\u017e v roce 1968 – font minimalizuje riziko z\u00e1m\u011bny tvarov\u011b podobn\u00fdch znak\u016f p\u0159i strojov\u00e9m \u010dten\u00ed. Bohu\u017eel ale na \u00fakor esteti\u010dnosti a \u010ditelnosti pro \u010dlov\u011bka.\u00a0<\/p>\t\t\t\t\t\t\t\t<\/div>\n\t\t\t\t<\/div>\n\t\t\t\t\t<\/div>\n\t\t<\/div>\n\t\t\t\t\t<\/div>\n\t\t<\/section>\n\t\t\t\t

    \n\t\t\t\t\t\t
    \n\t\t\t\t\t
    \n\t\t\t
    \n\t\t\t\t\t\t
    \n\t\t\t\t
    \n\t\t\t\t\t\t\t
    \n\t\t\t
    <\/div>\n\t\t<\/div>\n\t\t\t\t\t\t<\/div>\n\t\t\t\t<\/div>\n\t\t\t\t\t<\/div>\n\t\t<\/div>\n\t\t\t\t\t<\/div>\n\t\t<\/section>\n\t\t\t\t
    \n\t\t\t\t\t\t
    \n\t\t\t\t\t
    \n\t\t\t
    \n\t\t\t\t\t\t
    \n\t\t\t\t
    \n\t\t\t\t\t\t\t\t\t

    Jak na to? Odpov\u011bd\u00ed je neuronov\u00e1 s\u00ed\u0165.<\/h2>

    Na p\u0159\u00edkladech v\u00fd\u0161e jsme si uk\u00e1zali, \u017ee probl\u00e9m\u016f p\u0159i \u010dten\u00ed textu n\u00e1s m\u016f\u017ee potkat hned cel\u00e1 \u0159ada. Za zm\u00ednku tak\u00e9 ur\u010dit\u011b stoj\u00ed \u010dasov\u00e1 n\u00e1ro\u010dnost, kter\u00e1 je spojena s nastaven\u00edm vhodn\u00fdch parametr\u016f funkc\u00ed pro vyhled\u00e1n\u00ed textu, jeho segmentaci a nakonec s p\u0159\u00edpadnou tvorbou modelu pro rozpozn\u00e1n\u00ed znak\u016f pou\u017eit\u00e9ho fontu. Nic v\u0161ak nen\u00ed nemo\u017en\u00e9 a postupn\u00fdm lad\u011bn\u00edm kombinace parametr\u016f s testov\u00e1n\u00edm na dostate\u010dn\u011b velk\u00e9 a reprezentativn\u00ed sad\u011b testovac\u00edch sn\u00edmk\u016f lze \u010dasto dos\u00e1hnout velmi stabiln\u00edch v\u00fdsledk\u016f.\u00a0<\/p>

    Ne\u0161lo by to n\u011bjak jednodu\u0161eji – a ve v\u00fdsledku t\u0159eba i levn\u011bji…?<\/h4><\/div>

    Na\u0161t\u011bst\u00ed \u0161lo. V r\u00e1mci v\u00fdvoje v oblasti um\u011bl\u00e9 inteligence doch\u00e1z\u00ed v posledn\u00ed dob\u011b k velmi rychl\u00e9mu pokroku a pozadu rozhodn\u011b nez\u016fst\u00e1v\u00e1 an\u00ed segment strojov\u00e9ho vid\u011bn\u00ed a zpracov\u00e1n\u00ed obrazu. Krom\u011b konven\u010dn\u00edch OCR funkc\u00ed uk\u00e1zan\u00fdch v\u00fd\u0161e nab\u00edz\u00ed SW pro zpracov\u00e1n\u00ed obrazu Zebra Aurora Vision<\/a> tak\u00e9 funkci OCR zalo\u017eenou na principu hlubok\u00fdch neuronov\u00fdch s\u00edt\u00ed. Zn\u00ed to slo\u017eit\u011b, ale pou\u017eit\u00ed je velmi jednoduch\u00e9. Tato jedna kouzeln\u00e1 funkce v sob\u011b integruje v podstat\u011b cel\u00fd proces segmentace a n\u00e1sledn\u00e9ho rozpozn\u00e1n\u00ed znak\u016f p\u0159i nastaven\u00ed pouze n\u011bkolika z\u00e1kladn\u00edch parametr\u016f – zejm\u00e9na velikosti znak\u016f a rozsahu o\u010dek\u00e1van\u00fdch znak\u016f v r\u00e1mci textu. Dokonce ani nutn\u011b nevad\u00ed pokud znaky m\u011bn\u00ed polaritu. Funkce toti\u017e umo\u017e\u0148uje pracovat i s jej\u00ed kombinovanou variantou. Neuronov\u00e1 s\u00ed\u0165 na pozad\u00ed funkce je nav\u00edc natr\u00e9nov\u00e1na na opravdu \u0161irokou \u0161k\u00e1lu font\u016f a v drtiv\u00e9 v\u011bt\u0161in\u011b p\u0159\u00edpad\u016f tak v\u0161e funguje doslova out-of-box, bez nutnosti adaptace na konkr\u00e9tn\u00ed pou\u017eit\u00fd font. Toto \u0159e\u0161en\u00ed tak nejen \u017ee p\u0159in\u00e1\u0161\u00ed dosud nev\u00eddanou robustnost rozpozn\u00e1n\u00ed textu, ale tak\u00e9 velmi v\u00fdrazn\u011b zkracuje dobu implementace, jeliko\u017e odpad\u00e1 \u010dasov\u011b n\u00e1ro\u010dn\u00e9 lad\u011bn\u00ed velk\u00e9ho mno\u017estv\u00ed parametr\u016f. To v\u0161e p\u0159i n\u00edzk\u00e9 v\u00fdpo\u010detn\u00ed n\u00e1ro\u010dnosti, p\u0159edur\u010duj\u00edc\u00ed jej\u00ed pou\u017eit\u00ed i v rychl\u00fdch procesech.<\/p>\t\t\t\t\t\t\t\t<\/div>\n\t\t\t\t<\/div>\n\t\t\t\t\t<\/div>\n\t\t<\/div>\n\t\t\t\t\t<\/div>\n\t\t<\/section>\n\t\t\t\t

    \n\t\t\t\t\t\t
    \n\t\t\t\t\t
    \n\t\t\t
    \n\t\t\t\t\t\t
    \n\t\t\t\t
    \n\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\"\"\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t<\/div>\n\t\t\t\t<\/div>\n\t\t\t\t\t<\/div>\n\t\t<\/div>\n\t\t\t\t\t<\/div>\n\t\t<\/section>\n\t\t\t\t
    \n\t\t\t\t\t\t
    \n\t\t\t\t\t
    \n\t\t\t
    \n\t\t\t\t\t\t
    \n\t\t\t\t
    \n\t\t\t\t\t\t\t
    \n\t\t\t
    <\/div>\n\t\t<\/div>\n\t\t\t\t\t\t<\/div>\n\t\t\t\t<\/div>\n\t\t\t\t\t<\/div>\n\t\t<\/div>\n\t\t\t\t\t<\/div>\n\t\t<\/section>\n\t\t\t\t
    \n\t\t\t\t\t\t
    \n\t\t\t\t\t
    \n\t\t\t
    \n\t\t\t\t\t\t
    \n\t\t\t\t
    \n\t\t\t\t\t\t\t\t\t
    Zaujaly v\u00e1s mo\u017enosti pokro\u010dil\u00e9ho rozpozn\u00e1n\u00ed textu a cht\u011bli byste se dozv\u011bd\u011bt v\u00edce? Nev\u00e1hejte n\u00e1s kontaktovat!<\/a><\/h6>\t\t\t\t\t\t\t\t<\/div>\n\t\t\t\t<\/div>\n\t\t\t\t\t<\/div>\n\t\t<\/div>\n\t\t\t\t\t<\/div>\n\t\t<\/section>\n\t\t\t\t<\/div>","protected":false},"excerpt":{"rendered":"

    OCR: \u010cten\u00ed textu v pr\u016fmyslov\u00fdch aplikac\u00edch pomoc\u00ed deep learning neuronov\u00fdch s\u00edt\u00ed P\u00e1r slov \u00favodem… I p\u0159es to, \u017ee se v pr\u016fmyslu k […]<\/p>","protected":false},"author":1,"featured_media":2966,"comment_status":"closed","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[1],"tags":[],"class_list":["post-2928","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-uncategorized"],"_links":{"self":[{"href":"https:\/\/visimatic.cz\/en\/wp-json\/wp\/v2\/posts\/2928","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/visimatic.cz\/en\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/visimatic.cz\/en\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/visimatic.cz\/en\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/visimatic.cz\/en\/wp-json\/wp\/v2\/comments?post=2928"}],"version-history":[{"count":47,"href":"https:\/\/visimatic.cz\/en\/wp-json\/wp\/v2\/posts\/2928\/revisions"}],"predecessor-version":[{"id":3001,"href":"https:\/\/visimatic.cz\/en\/wp-json\/wp\/v2\/posts\/2928\/revisions\/3001"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/visimatic.cz\/en\/wp-json\/wp\/v2\/media\/2966"}],"wp:attachment":[{"href":"https:\/\/visimatic.cz\/en\/wp-json\/wp\/v2\/media?parent=2928"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/visimatic.cz\/en\/wp-json\/wp\/v2\/categories?post=2928"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/visimatic.cz\/en\/wp-json\/wp\/v2\/tags?post=2928"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}