Abacwaningi abavela elabhorethri iNyuvesi yaseChicago yakha isethi yamathuluzi ngokuqaliswa ukuhlanekezelwa kwezithombe, ukuvimbela ukusetshenziswa kwazo ekuqeqesheni ukuqashelwa kobuso kanye nezinhlelo zokuhlonza abasebenzisi. Izinguquko zephikseli zenziwa esithombeni, ezingabonakali uma zibukwa abantu, kodwa ziholela ekwakhekeni kwamamodeli angalungile uma esetshenziselwa ukuqeqesha amasistimu okufunda omshini. Ikhodi yekhithi yamathuluzi ibhalwe kuPython futhi ngaphansi kwelayisensi ye-BSD. Imihlangano yeLinux, macOS neWindows.
Ukucubungula izithombe ngesisetshenziswa esihlongozwayo ngaphambi kokushicilela kumanethiwekhi omphakathi nezinye izinkundla zomphakathi kukuvumela ukuthi uvikele umsebenzisi ekusebenziseni idatha yezithombe njengomthombo wokuqeqesha amasistimu okubona ubuso. I-algorithm ehlongozwayo inikeza isivikelo ku-95% wemizamo yokuqaphela ubuso (yeMicrosoft Azure recognition API, Amazon Rekognition and Face++, ukusebenza kahle kokuvikela kungu-100%). Ngaphezu kwalokho, noma ngabe esikhathini esizayo izithombe zangempela, ezingacutshungulwanga yi-utility, zisetshenziswa kumodeli esivele iqeqeshwe kusetshenziswa izinguqulo ezihlanekezelwe zezithombe, izinga lokwehluleka ekuqashelweni lihlala lifana futhi okungenani lingama-80%.
Indlela isekelwe kulokho "izibonelo eziphikisanayo", ingqikithi yayo ukuthi izinguquko ezincane kudatha yokufaka zingaholela ezinguqukweni ezimangalisayo kumqondo wokuhlukanisa. Njengamanje, isenzakalo "sezibonelo eziphikisanayo" singenye yezinkinga eziyinhloko ezingaxazululiwe ezinhlelweni zokufunda zomshini. Ngokuzayo, isizukulwane esisha sezinhlelo zokufunda zomshini kulindeleke ukuthi sivele esingenayo le nkinga, kodwa lezi zinhlelo zizodinga izinguquko ezibalulekile ekwakhiweni kwezakhiwo kanye nendlela yamamodeli wokwakha.
Ukucubungula izithombe kwehla ekungezeni inhlanganisela yamaphikseli (amaqoqo) esithombeni, abonwa ngama-algorithms okufunda omshini ajulile njengamaphethini esici sento efanekisiwe futhi aholela ekuhlanekeni kwezici ezisetshenziselwa ukuhlukanisa. Izinguquko ezinjalo azihlukani nesethi evamile futhi kunzima kakhulu ukuzibona nokuzikhipha. Ngisho nezithombe zangempela nezilungisiwe, kunzima ukunquma ukuthi iyiphi eyangempela nokuthi iyiphi inguqulo eguquliwe.
Ukuhlanekezela okwethulwe kubonisa ukumelana okuphezulu ekudalweni kwezinyathelo ezihloselwe ukuhlonza izithombe ezephula ukwakhiwa okulungile kwamamodeli okufunda omshini. Kubandakanya izindlela ezisuselwe ekufiphaleni, ekungezeni umsindo, noma ukusebenzisa izihlungi esithombeni ukuze ucindezele izinhlanganisela zamaphikseli azisebenzi. Inkinga iwukuthi uma izihlungi zisetshenziswa, ukunemba kwezigaba kwehla ngokushesha kakhulu kunokutholakala kwamaphethini we-pixel, futhi ezingeni lapho ukuhlanekezela kucindezelwe, izinga lokuqashelwa ngeke lisabhekwa njengelamukelekayo.
Kuyaphawulwa ukuthi, njengobunye ubuchwepheshe obuningi bokuvikela ubumfihlo, indlela ehlongozwayo ingasetshenziswa hhayi kuphela ukulwa nokusetshenziswa okungagunyaziwe kwezithombe zomphakathi ezinhlelweni zokuqaphela, kodwa futhi njengethuluzi lokufihla abahlaseli. Abacwaningi bakholelwa ukuthi izinkinga zokuqashelwa zingase zithinte ngokuyinhloko izinsizakalo zezinkampani zangaphandle eziqoqa ulwazi ngokungalawuleki nangaphandle kwemvume yokuqeqesha amamodeli abo (ngokwesibonelo, isevisi ye-Clearview.ai inikeza isizindalwazi sokubona ubuso, cishe izithombe eziyizigidi eziyizinkulungwane ezingu-3 ezivela ezinkundleni zokuxhumana zibaliwe). Uma manje amaqoqo ezinsizakalo ezinjalo aqukethe izithombe ezinokwethenjelwa kakhulu, ngakho-ke ngokusetshenziswa okusebenzayo kwe-Fawkes, ngokuhamba kwesikhathi, isethi yezithombe ezihlanekezelwe izoba nkulu futhi imodeli izobabheka njengento ebaluleke kakhulu ekuhlukaniseni. Izinhlelo zokuqashelwa ze-ejensi yezobunhloli, amamodeli azo akhiwe ngesisekelo semithombo ethembekile, azothinteka kancane ngamathuluzi ashicilelwe.
Phakathi kwentuthuko engokoqobo esondelene nenhloso, singaphawula iphrojekthi , asathuthuka ukwengeza ezithombeni , ukuvimbela ukuhlukaniswa okufanele ngezinhlelo zokufunda zomshini. Ikhodi ye-Adversaria yekhamera ku-GitHub ngaphansi kwelayisensi ye-EPL. Enye iphrojekthi ihlose ukuvimba ukuqashelwa ngamakhamera okuqapha ngokudala amajazi emvula anephethini ekhethekile, izikibha, amajezi, amakepisi, amaphosta noma izigqoko.
Source: opennet.ru