Isu lokuhlanekezela izithombe ngobuqili ukuphazamisa amasistimu okubona ubuso

Abacwaningi abavela elabhorethri Sand Inyuvesi yaseChicago yakha isethi yamathuluzi fawkes ngokuqaliswa indlela ukuhlanekezelwa kwezithombe, ukuvimbela ukusetshenziswa kwazo ekuqeqesheni ukuqashelwa kobuso kanye nezinhlelo zokuhlonza abasebenzisi. Izinguquko zephikseli zethulwa esithombeni, ezingabonakali kubantu kodwa ziholela ekwakhekeni kwamamodeli angalungile uma esetshenziselwa ukuqeqesha amasistimu okufunda omshini. Ikhodi yekhithi yamathuluzi ibhalwe kuPython futhi eshicilelwe ngaphansi kwelayisensi ye-BSD. Imihlangano ilungisiwe yeLinux, macOS neWindows.

Ukucubungula izithombe ngesisetshenziswa esihlongozwayo ngaphambi kokuzishicilela ezinkundleni zokuxhumana nakwezinye izinkundla zomphakathi kuvikela abasebenzisi ekusebenziseni lezi zithombe njengomthombo wokuqeqeshwa wamasistimu okubona ubuso. I-algorithm ehlongozwayo ihlinzeka ngokuvikeleka ku-95% wemizamo yokuqashelwa kobuso (ye-Microsoft Azure, Amazon Rekognition, kanye ne-Face++ recognition APIs, ukusebenza kahle kokuvikela kungu-100%). Ngaphezu kwalokho, ngisho noma izithombe zangempela, ezingacutshungulwanga kamuva zisetshenziswa kumodeli eyaqeqeshwa kusetshenziswa izinguqulo ezihlanekezelwe zezithombe, izinga lokwehluleka ukuqashelwa lihlala okungenani lingu-80%.

Indlela isuselwe kwinto "yezibonelo eziphikisanayo," ethi izinguquko ezincane kudatha yokufaka zingaholela ezinguqukweni ezimangalisayo kumqondo wokuhlukanisa. Okwamanje, okwenzeka "kwezibonelo eziphikisanayo" kungenye yezinkinga ezinkulu ezingaxazululiwe kumasistimu okufunda emishini. Ngokuzayo, isizukulwane esisha sezinhlelo zokufunda zemishini esingenalo mphumela kulindeleke ukuthi sivele, kodwa lezi zinhlelo zizodinga izinguquko eziphawulekayo ekwakhiweni kwazo kanye nendlela yokwakha imodeli.

Ukucutshungulwa kwesithombe kuhlanganisa ukungeza inhlanganisela yamaphikseli (amaqoqo) esithombeni. Lawa maqoqo atolikwa ngama-algorithms okufunda ajulile njengamaphethini esici sento eboniswayo, ehlanekezela izici ezisetshenziselwa ukuhlukanisa. Izinguquko ezinjalo azihlukaniseki kusethi evamile futhi kunzima kakhulu ukuzibona nokuzikhipha. Ngisho nakuzo zombili izithombe zangempela nezilungisiwe, kunzima ukuhlukanisa phakathi kwenguqulo yokuqala neshintshiwe.

Ukuhlanekezela okwethulwe kubonisa ukumelana okuphezulu nezinyathelo ezihloselwe ukuhlonza izithombe eziphazamisa ukwakhiwa okulungile kwamamodeli okufunda omshini. Phakathi kwezinye izinto, izindlela ezisuselwe ekufiphaleni, ekungezeni umsindo, noma ukusebenzisa izihlungi esithombeni ukucindezela amaphethini wamaphikseli azisebenzi. Inkinga ukuthi uma usebenzisa izihlungi, ukunemba kwezigaba kwehla ngokushesha kakhulu kunokutholakala kwamaphethini e-pixel, futhi uma ukuhlanekezela sekucindezelwe, izinga lokuqaphela 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 ngokuyinhloko zingathinta izinsizakalo zezinkampani zangaphandle ezingalawuleki futhi ngaphandle kwemvume eziqoqa idatha ukuze ziqeqeshe amamodeli azo (ngokwesibonelo, isevisi ye-Clearview.ai inikeza isizindalwazi sokubona ubuso, eyakhelwe (Lokhu kusekelwe ekukhombeni cishe izithombe eziyizigidi eziyizinkulungwane ezingu-3 ezinkundleni zokuxhumana.) Nakuba amaqoqo ezinsizakalo ezinjalo okwamanje aqukethe izithombe ezinokwethenjelwa kakhulu, ngokusetshenziswa okusebenzayo kwe-Fawkes, ngokuhamba kwesikhathi, isethi yezithombe ezihlanekezelwe izokwanda, futhi imodeli izozibeka phambili ukuze zihlukaniswe. Ikhithi yamathuluzi eshicilelwe izoba nomthelela omncane ezinhlelweni zokuqashelwa ze-ejensi yezobunhloli, amamodeli azo asekelwe emithonjeni ethembekile.

Phakathi kwentuthuko engokoqobo esondelene nenhloso, umuntu angabona iphrojekthi I-Adversaria yekhamera, asathuthuka uhlelo lokusebenza lweselula ukwengeza ezithombeni Umsindo wePerlin, okuphazamisa ukuhlukaniswa okufanele ngamasistimu okufunda omshini. Ikhodi ye-Adversaria yekhamera iyatholakala ku-GitHub ngaphansi kwelayisensi ye-EPL. Enye iphrojekthi Ijazi elingabonakali ihlose ukuvimba ukuqashelwa ngamakhamera okuqapha ngokudala amajazi emvula anephethini ekhethekile, izikibha, amajezi, amakepisi, amaphosta noma izigqoko.

Source: opennet.ru