Uhlelo olusha lokuqaphela umbhalo we-EasyOCR

Iphrojekthi I-EasyOCR Kuthuthukiswa uhlelo olusha lokubona umbhalo obonakalayo olusekela izilimi ezingaphezu kuka-40, okuhlanganisa isiNgisi, isiJalimane, isiFulentshi, isiJapane, isiShayina, isiKorea, isi-Uzbek, isi-Azerbaijani nesiLithuanian. Izilimi ezisekelwe kuCyrillic azikasekelwa, kodwa ziyengezwa ohlwini lwezinhlelo. Ikhodi ibhalwe ngePython kusetshenziswa uhlaka I-PyTorch и isatshalaliswa ngu ilayisensi ngaphansi kwe-Apache 2.0. Okokulayisha ahlinzekwa amamodeli enziwe ngomumo wezilimi asuselwa ku-alfabhethi yesiLatini nama-hieroglyphs.

Izindlela zokufunda ngomshini zisetshenziselwa ukukhomba nokubona umbhalo osesithombeni. I-algorithm yokufunda yomshini isetshenziselwa ukukhomba umbhalo UBUCIKO (Ukuqwashisa Ngezinhlamvu-Isifunda Sombhalo) ku ukuqaliswa ye-PyTorch, ekwazi ukugqamisa umbhalo ezintweni ezingahleliwe, okuhlanganisa amalebula, izimpawu zolwazi nezimpawu zomgwaqo. Inethiwekhi ye-convolutional recurrent neural isetshenziswa ukubona ukulandelana kwezinhlamvu I-CRNN (I-Convolutional Recurrent Neural Network, inhlanganisela ye-DCNN ne-RNN) kanye ne-algorithm I-CTC BeamSearch I-CTC BeamSearch (I-Connectionist Temporal Classification) ukuze inqume ukuphuma kwenethiwekhi ye-neural ibe ukumelwa kombhalo.

Source: opennet.ru