Inkqubo entsha yokuqaphela umbhalo we-EasyOCR

Iprojekthi I-EasyOCR Inkqubo entsha yokuqaphela umbhalo we-optical iphuhliswa exhasa iilwimi ezingaphezu kwe-40, kuquka isiNgesi, isiJamani, isiFrentshi, isiJapan, isiTshayina, isiKorea, isiUzbek, isiAzerbaijani kunye nesiLithuanian. Iilwimi ezisekelwe kwiCyrillic azikaxhaswa, kodwa zongezwa kuluhlu lwezicwangciso. Ikhowudi ibhalwe kwiPython usebenzisa isakhelo I-PyTorch ΠΈ isasazwa ngu ilayisenisi phantsi kweApache 2.0. Yokulayisha zibonelelwe iimodeli esele zenziwe ngeelwimi ezisekwe kwialfabhethi yesiLatini kunye ne-hieroglyphs.

Iindlela zokufunda ngoomatshini zisetyenziselwa ukuchonga nokuqaphela isicatshulwa kumfanekiso. I-algorithm yokufunda ngomatshini isetyenziselwa ukuchonga isicatshulwa Craft (Ulwazi ngoMlinganiswa-kwiNgingqi kwiSibhalo) kwi ukuphunyezwa yePyTorch, ekwaziyo ukuqaqambisa okubhaliweyo kwizinto ezingafunekiyo, kuquka iilebhile, iimpawu zolwazi kunye neempawu zendlela. Inethiwekhi ye-convolutional recurrent neural isetyenziselwa ukuqaphela ulandelelwano lwabalinganiswa CRNN (I-Convolutional Recurrent Neural Network, indibaniselwano ye-DCNN kunye ne-RNN) kunye ne-algorithm CTC BeamSearch I-CTC BeamSearch (uHlelo lweThutyana loQhagamshelwano) ukucacisa imveliso yenethiwekhi ye-neural kumelo lombhalo.

umthombo: opennet.ru

Yongeza izimvo