Njira yosokoneza zithunzi mochenjera kuti zisokoneze machitidwe ozindikira nkhope

Ofufuza ochokera ku labotale SAND Yunivesite ya Chicago inapanga zida zothandizira zovuta ndi kukhazikitsa njira kusokoneza zithunzi, kulepheretsa kugwiritsa ntchito kwawo pophunzitsa kuzindikira nkhope ndi machitidwe ozindikiritsa ogwiritsa ntchito. Kusintha kwa pixel kumapangidwira chithunzicho, chomwe sichiwoneka pamene anthu amachiwona, koma chimapangitsa kuti pakhale zitsanzo zolakwika zomwe zimagwiritsidwa ntchito pophunzitsa makina ophunzirira makina. Khodi ya zida zidalembedwa mu Python ndi losindikizidwa pansi pa layisensi ya BSD. Misonkhano kukonzekera kwa Linux, macOS ndi Windows.

Kukonza zithunzi ndi zomwe mukufuna kugwiritsa ntchito musanasindikizidwe pa malo ochezera a pa Intaneti ndi malo ena ochezera anthu kumakupatsani mwayi woteteza wogwiritsa ntchito zithunzi ngati gwero lophunzitsira machitidwe ozindikira nkhope. Ma algorithm omwe akufunsidwa amapereka chitetezo ku 95% ya kuyesa kuzindikira nkhope (kwa Microsoft Azure recognition API, Amazon Rekognition ndi Face ++, chitetezo chokwanira ndi 100%). Komanso, ngakhale m'tsogolomu zithunzi zoyambirira, zosagwiritsidwa ntchito ndi zothandizira, zimagwiritsidwa ntchito mu chitsanzo chomwe chaphunzitsidwa kale pogwiritsa ntchito zithunzi zopotoka, kuchuluka kwa kulephera kuzindikira kumakhalabe komweko ndipo ndi osachepera 80%.

Njirayi imachokera ku zochitika za "zitsanzo zotsutsana", zomwe zimachititsa kuti kusintha kwakung'ono kwa deta yolowetsamo kungayambitse kusintha kwakukulu pamalingaliro amagulu. Pakalipano, chodabwitsa cha "zitsanzo zotsutsana" ndi chimodzi mwa zovuta zomwe sizinathetsedwe mu makina ophunzirira makina. M'tsogolomu, mbadwo watsopano wamakina ophunzirira makina ukuyembekezeka kuwonekera womwe ulibe vuto ili, koma machitidwewa adzafuna kusintha kwakukulu kwa zomangamanga ndi njira yopangira zitsanzo zomanga.

Kukonza zithunzi kumatsikira pakuwonjezera ma pixel (magulu) pachithunzichi, zomwe zimazindikirika ndi makina ozama ophunzirira makina ngati mawonekedwe a chinthu chojambulidwa ndikupangitsa kusokoneza zomwe zimagwiritsidwa ntchito pogawa. Zosintha zotere sizimasiyana ndi zonse ndipo zimakhala zovuta kuzizindikira ndikuzichotsa. Ngakhale ndi zithunzi zoyambirira ndi zosinthidwa, n'zovuta kudziwa chomwe chiri choyambirira komanso chomwe chiri chosinthidwa.

Zopotoka zomwe zatulutsidwa zikuwonetsa kukana kwakukulu pakupanga njira zotsutsana ndi zomwe zikufuna kuzindikira zithunzi zomwe zimaphwanya kapangidwe koyenera kamitundu yophunzirira makina. Kuphatikizirapo njira zotengera kusawoneka bwino, kuwonjezera phokoso, kapena kugwiritsa ntchito zosefera pachithunzichi kuti kupondereza kuphatikiza kwa pixel sizothandiza. Vuto ndiloti zosefera zikagwiritsidwa ntchito, kulondola kwamagulu kumatsika mwachangu kuposa kuzindikirika kwa mawonekedwe a pixel, ndipo pamlingo pomwe zosokoneza zimaponderezedwa, mulingo wozindikirika sungakhalenso wovomerezeka.

Zimadziwika kuti, monga njira zina zambiri zamakono zotetezera zinsinsi, njira yomwe ikuperekedwayo ingagwiritsidwe ntchito osati kulimbana ndi kugwiritsidwa ntchito kosaloledwa kwa zithunzi za anthu m'machitidwe ozindikiritsa, komanso ngati chida chobisala omwe akuukira. Ochita kafukufuku amakhulupirira kuti mavuto ozindikirika angakhudze kwambiri ntchito zamagulu ena omwe amasonkhanitsa zambiri mosalamulirika komanso popanda chilolezo chophunzitsa zitsanzo zawo (mwachitsanzo, ntchito ya Clearview.ai imapereka database yozindikiritsa nkhope, anamanga pafupifupi zithunzi 3 biliyoni zochokera kumalo ochezera a pa Intaneti zalembedwa). Ngati tsopano zosonkhanitsira zantchito zotere zili ndi zithunzi zodalirika, ndiye kuti ndikugwiritsa ntchito Fawkes mwachangu, pakapita nthawi, zithunzi zopotoka zidzakhala zazikulu ndipo chitsanzocho chiziwona ngati chofunikira kwambiri pagulu. Njira zozindikiritsa za mabungwe anzeru, zitsanzo zomwe zimamangidwa pamaziko odalirika, sizidzakhudzidwa kwambiri ndi zida zofalitsidwa.

Pakati pa zochitika zothandiza zomwe zili pafupi ndi cholinga, tikhoza kuzindikira polojekitiyi Kamera Adversaria, kukula pulogalamu yam'manja kuwonjezera ku zithunzi Perlin phokoso, kuletsa kugawidwa koyenera ndi makina ophunzirira makina. Kamera Adversaria Code zilipo pa GitHub pansi pa layisensi ya EPL. Ntchito ina Chovala chosaoneka ikufuna kuletsa kuzindikirika ndi makamera oyang'anira popanga majeketi apadera amvula, T-shirts, majuzi, zipewa, zikwangwani kapena zipewa.

Source: opennet.ru