Algorithms na sarrafa bayanai kamar hanyoyin sadarwa na jijiyoyi sun dauki duniya da hadari. Ci gaban su shine saboda dalilai da yawa, ciki har da kayan aiki masu arha da ƙarfi da adadi mai yawa na bayanai. Hanyoyin sadarwa na jijiyoyi a halin yanzu suna kan gaba ga duk abin da ke da alaka da ayyukan "fahimi" kamar gane hoto, fahimtar harshen yanayi, da dai sauransu. Amma bai kamata a iyakance su ga irin waɗannan ayyuka ba. Wannan labarin yana magana ne game da yadda ake damfara hotuna ta amfani da hanyoyin sadarwa na jijiyoyi, ta amfani da ragowar koyo. Hanyar da aka gabatar a cikin labarin ya fi sauri kuma mafi kyau fiye da daidaitattun codecs. Tsare-tsare, daidaito da kuma, ba shakka, tebur tare da gwaje-gwaje a ƙarƙashin yanke.
Wannan labarin ya dogara ne akan aiki. Ana ɗauka cewa kun saba da hanyoyin sadarwa na jijiyoyi da ra'ayoyinsu. rikice-rikice и aikin hasara.
Menene matsawar hoto kuma ta yaya yake aiki?
Matsa hoto shine tsarin canza hoto ta yadda zai ɗauki ƙasa da sarari. Ajiye hotuna kawai zai ɗauki sarari da yawa, shi ya sa ake samun codecs kamar JPEG da PNG waɗanda ke nufin rage girman ainihin hoton.
Kamar yadda ka sani, akwai nau'ikan damfara hoto iri biyu: babu asara и tare da hasara. Kamar yadda sunayen ke ba da shawara, matsawa mara asara na iya riƙe ainihin bayanan hoton, yayin da matsi na asarar wasu bayanai yayin matsawa. Misali, JPG sune algorithms masu asara [kimanin. fassara - Ainihin, kar mu manta game da JPEG mara hasara], kuma PNG algorithm ne mara asara.
Kwatanta rashin asara da matsi
Lura cewa akwai abubuwa masu toshewa da yawa a cikin hoton da ke hannun dama. Wannan bayanin batattu ne. Maƙwabtan pixels masu launi iri ɗaya ana matsa su azaman yanki ɗaya don adana sarari, amma bayanai game da ainihin pixels sun ɓace. Tabbas, algorithms da aka yi amfani da su a cikin JPEG, PNG, da sauransu. codecs sun fi rikitarwa, amma wannan kyakkyawan misali ne mai hankali na matsi mai asara. Matsi mara hasara yana da kyau, amma fayilolin da aka matsa marasa asara suna ɗaukar sararin diski mai yawa. Akwai ingantattun hanyoyi don damfara hotuna ba tare da rasa bayanai da yawa ba, amma suna da sannu a hankali kuma da yawa suna amfani da hanyoyin maimaitawa. Wannan yana nufin cewa ba za a iya gudanar da su a layi daya akan yawancin CPU ko GPU ba. Wannan ƙayyadaddun ya sa su gaba ɗaya ba su da amfani a amfanin yau da kullun.
Shigar da hanyar sadarwa ta Convolutional Neural
Idan wani abu yana buƙatar ƙididdigewa kuma ƙididdiga na iya zama kusan, ƙara . Marubutan sun yi amfani da daidaitaccen madaidaicin hanyar sadarwa na juzu'i don inganta damfara hoto. Hanyar da aka gabatar ba kawai tana yin daidai da mafi kyawun mafita ba (idan ba mafi kyau ba), kuma yana iya amfani da lissafin layi ɗaya, wanda ke haifar da karuwa mai girma a cikin sauri. Dalili kuwa shi ne Convolutional Neural Networks (CNNs) sun ƙware sosai wajen fitar da bayanan sararin samaniya daga hotuna, waɗanda sai a gabatar da su a cikin mafi ƙanƙanta tsari (misali, kawai “mahimmanci” na hoton an adana su). Mawallafa sun so su yi amfani da wannan damar CNN don mafi kyawun wakilcin hotuna.
gine
Marubutan sun ba da shawarar hanyar sadarwa biyu. Cibiyar sadarwa ta farko tana ɗaukar hoto azaman shigarwa kuma tana haifar da ƙaramin wakilci (ComCNN). Ana sarrafa fitar da wannan hanyar sadarwa ta daidaitaccen codec (misali JPEG). Bayan an sarrafa shi ta hanyar codec, hoton yana wucewa zuwa hanyar sadarwa ta biyu, wanda ke "gyara" hoton daga codec a ƙoƙarin dawo da ainihin hoton. Marubutan sun ba wa wannan hanyar sadarwa suna CNN Reconstructive CNN (RecCNN). Kamar GANs, duka cibiyoyin sadarwa ana horar da su akai-akai.
ComCNN Compact an wuce shi zuwa daidaitaccen codec
RecCNN. An haɓaka fitarwar ComCNN kuma an ciyar da shi zuwa RecCNN, wanda zai yi ƙoƙarin koyon sauran
Ana haɓaka fitarwar codec sama sannan a wuce zuwa RecCNN. RecCNN zai yi ƙoƙarin sanya hoton kusa da ainihin yadda zai yiwu.
Ƙarshe-zuwa-ƙarshen tsarin matsa hoto. Co(.) algorithm ne na matsa hoto. Marubutan sun yi amfani da JPEG, JPEG2000 da BPG
Menene saura?
Za a iya tunanin ragowar a matsayin mataki na gaba don "inganta" hoton da codec ke yankewa. Samun "bayanai" da yawa game da duniya, cibiyar sadarwar jijiyoyi na iya yin yanke shawara na hankali game da abin da za a gyara. Wannan ra'ayin ya dogara ne akan , karanta cikakken bayani game da abin da za ku iya .
Ayyukan hasara
Ana amfani da ayyukan asara guda biyu saboda muna da cibiyoyin sadarwa guda biyu. Na farkon waɗannan, ComCNN, ana yiwa lakabi da L1 kuma an siffanta shi kamar haka:
Ayyukan hasara don ComCNN
Bayani
Wannan lissafin yana iya zama kamar rikitarwa, amma ainihin ma'auni ne (kuskuren murabba'i tushen tushen) MSE. ||² yana nufin ƙa'idar vector da suka haɗa.
Daidaitawa 1.1
Cr yana nuna fitowar ComCNN. θ yana nuna ƙwarewar koyan sigogin ComCNN, XK shine hoton shigarwa
Daidaitawa 1.2
Re() yana nufin RecCNN. Wannan lissafin kawai yana isar da ma'anar lissafin 1.1 zuwa RecCNN. θ yana nuna sigogin horarwa na RecCNN (hati a saman yana nufin an gyara sigogi).
Ma'anar Intuitive
Equation 1.0 zai sa ComCNN ya canza ma'auninsa ta yadda idan aka sake yin shi da RecCNN, hoton ƙarshe ya yi kama da mai yiwuwa ga hoton shigarwa. An bayyana aikin asarar RecCNN na biyu kamar haka:
Daidaitawa 2.0
Bayani
Bugu da ƙari, aikin na iya yin kama da rikitarwa, amma wannan ga mafi yawan ɓangaren aikin asarar cibiyar sadarwa ne (MSE).
Daidaitawa 2.1
Co() yana nufin fitarwa na codec, x tare da hula a saman yana nufin fitarwar ComCNN. θ2 sune sigogin horarwa na RecCNN, res() shine kawai RecCNN ta saura fitarwa. Yana da kyau a lura cewa an horar da RecCNN akan bambanci tsakanin Co () da hoton shigarwa, amma ba akan hoton shigarwa ba.
Ma'anar Intuitive
Matsakaicin 2.0 zai sa RecCNN ya canza ma'auninsa domin abin da ake fitarwa yayi kama da kama da hoton shigarwa.
Tsarin horo
Ana horar da samfura akai-akai, kamar . Ana daidaita ma'auni na samfurin farko yayin da ake sabunta ma'auni na na biyu, sa'an nan kuma an daidaita ma'aunin samfurin na biyu yayin da ake horar da samfurin farko.
Gwaje-gwaje
Marubutan sun kwatanta hanyarsu tare da hanyoyin da ake da su, gami da codecs masu sauƙi. Hanyar su tana aiki mafi kyau fiye da wasu yayin da suke riƙe babban gudu akan kayan aikin da suka dace. Bugu da ƙari, marubutan sun yi ƙoƙari su yi amfani da ɗaya kawai daga cikin hanyoyin sadarwa guda biyu kuma sun lura da raguwar aiki.
Kwatanta Ma'anar Daidaita Tsari (SSIM). Babban dabi'u suna nuna mafi kyawun kamanni da asali. Nau'in ƙarfin hali yana nuna sakamakon aikin marubuta
ƙarshe
Mun kalli sabuwar hanyar da za a yi amfani da zurfin ilmantarwa don matsawa hoto, kuma mun yi magana game da yiwuwar yin amfani da hanyoyin sadarwa na jijiyoyi a cikin ayyuka fiye da ayyukan "jama'a" irin su rarraba hoto da sarrafa harshe. Wannan hanyar ba wai kawai tana ƙasa da buƙatun zamani ba, amma kuma tana ba ku damar aiwatar da hotuna da sauri.
Koyon hanyoyin sadarwa na jijiyoyi ya zama mai sauƙi, saboda mun yi lambar talla musamman don Habravchan HABR, bada ƙarin rangwame 10% zuwa rangwamen da aka nuna akan banner.
Ƙarin darussa
Fitattun Labarai
source: www.habr.com