Ma algorithms oyendetsedwa ndi data monga ma neural network atenga dziko lapansi movutikira. Kukula kwawo kumayendetsedwa ndi zifukwa zingapo, kuphatikiza zida zotsika mtengo komanso zamphamvu komanso kuchuluka kwa data. Ma Neural network ali patsogolo pa chilichonse chokhudzana ndi ntchito "zachidziwitso" monga kuzindikira zithunzi, kumvetsetsa chilankhulo, ndi zina. Koma sayenera kungokhala pa ntchito zoterozo. Nkhaniyi ikufotokoza njira yopondereza zithunzi pogwiritsa ntchito ma neural network pogwiritsa ntchito kuphunzira kotsalira. Njira yomwe yafotokozedwa m'nkhaniyi imagwira ntchito mwachangu komanso bwino kuposa ma codec wamba. Schemes, equations ndipo, ndithudi, tebulo ndi mayesero pansi odulidwa.
Nkhaniyi yachokera pa ntchito. Zimaganiziridwa kuti mumadziwa ma neural network ndi malingaliro awo convolution ΠΈ ntchito yotaya.
Kodi kupanikizana kwa zithunzi ndi mitundu yanji komwe kumabwera?
Kuponderezana kwa zithunzi ndi njira yosinthira chithunzi kuti chitenge malo ochepa. Kungosunga zithunzi kungatenge malo ambiri, kotero pali ma codec monga JPEG ndi PNG omwe cholinga chake ndi kuchepetsa kukula kwa chithunzi choyambirira.
Monga mukudziwa, pali mitundu iwiri ya kupanikizana kwazithunzi: palibe kutaya ΠΈ ndi zotayika. Monga mayina amanenera, kukanikiza kopanda kutaya kumatha kuyambiranso deta yachifaniziro, pomwe kupanikizana kotayika kumataya data ina panthawi ya psinjika. mwachitsanzo, JPG ndi ma aligorivimu otayika [pafupifupi. kumasulira - makamaka, tisaiwalenso za JPEG yopanda kutaya], ndipo PNG ndi algorithm yopanda pake.
Kufananiza kupsinjika kosataya komanso kotayika
Zindikirani kuti chithunzi chakumanja chili ndi zinthu zambiri za blocky. Izi ndi zotayika. Ma pixel oyandikana amitundu yofananira amapanikizidwa ngati gawo limodzi kuti asunge malo, koma zambiri zama pixel enieni zimatayika. Zoonadi, ma aligorivimu omwe amagwiritsidwa ntchito mu JPEG, PNG, etc. Kuphatikizika kopanda kutaya ndikwabwino, koma mafayilo oponderezedwa osataya amatenga malo ambiri a disk. Pali njira zabwino zosindikizira zithunzi popanda kutaya zambiri, koma ndizochepa kwambiri ndipo ambiri amagwiritsa ntchito njira zobwerezabwereza. Izi zikutanthauza kuti sizingayendetsedwe mofanana pamagulu angapo a CPU kapena GPU. Izi zimawapangitsa kukhala osatheka kugwiritsa ntchito tsiku ndi tsiku.
Convolutional Neural Network Input
Ngati chinachake chikufunika kuwerengedwa ndipo mawerengedwe angakhale oyerekezera, onjezani . Olembawo adagwiritsa ntchito njira yolumikizira neural network kuti apititse patsogolo kupsinjika kwa zithunzi. Njira yoperekedwayo sikuti imangochita molingana ndi njira zabwino kwambiri (ngati sizili bwino), imathanso kugwiritsa ntchito computing yofananira, zomwe zimapangitsa kuti liwiro liwonjezeke kwambiri. Chifukwa chake ndi chakuti ma convolutional neural networks (CNNs) ndiabwino kwambiri pochotsa zidziwitso zapamalo kuchokera pazithunzi, zomwe zimayimiriridwa mwanjira yophatikizika (mwachitsanzo, "zofunikira" zokha zachithunzizo zimasungidwa). Olembawo ankafuna kugwiritsa ntchito mawonekedwe a CNN kuti awonetse bwino zithunzi.
zomangamanga
Olembawo adapereka maukonde apawiri. Netiweki yoyamba imatenga chithunzi ngati cholowetsa ndikupanga choyimira chophatikizika (ComCNN). Kutulutsa kwa netiwekiyi kumakonzedwa ndi codec yokhazikika (monga JPEG). Mukakonzedwa ndi codec, chithunzicho chimatumizidwa ku netiweki yachiwiri, yomwe "imawongolera" chithunzicho kuchokera ku codec poyesa kubwezera chithunzi choyambirira. Olembawo adatcha netiweki iyi CNN yomanganso (RecCNN). Monga ma GAN, maukonde onsewa amaphunzitsidwa mobwerezabwereza.
Chiwonetsero cha ComCNN Compact chasamutsidwa ku codec wamba
RecCNN. Kutulutsa kwa ComCNN kumakwezedwa ndikudyetsedwa ku RecCNN, yomwe idzayesa kuphunzira zotsalira
Kutulutsa kwa codec kumakwezedwa ndikuperekedwa ku RecCNN. RecCNN idzayesa kutulutsa chithunzi chomwe chili chofanana ndi choyambirira momwe kungathekere.
Kumapeto-pa-mapeto kwa chithunzithunzi chophatikizika. Co(.) ndi njira yopondereza zithunzi. Olembawo adagwiritsa ntchito JPEG, JPEG2000 ndi BPG
Chotsalira ndi chiyani?
Chotsaliracho chikhoza kuganiziridwa ngati sitepe yokonzekera "kupititsa patsogolo" chithunzi chomwe chikusinthidwa ndi codec. Ndi "chidziwitso" chochuluka chokhudza dziko lapansi, neural network ikhoza kupanga zisankho zachidziwitso pa zomwe angakonze. Lingaliro ili lachokera , werengani tsatanetsatane wa zomwe mungathe .
Kutaya ntchito
Ntchito ziwiri zotayika zimagwiritsidwa ntchito chifukwa tili ndi ma neural network awiri. Yoyamba mwa izi, ComCNN, imatchedwa L1 ndipo imafotokozedwa motere:
Kutaya ntchito kwa ComCNN
Kufotokozera
Equation iyi ikhoza kuwoneka yovuta, koma ndiyokhazikika (kutanthauza cholakwika cha squared) MSE. ||Β² amatanthauza chizolowezi cha vector yomwe amatsekera.
Equation 1.1
Cr amatanthauza kutulutsa kwa ComCNN. ΞΈ amatanthauza kuphunzitsidwa kwa magawo a ComCNN, XK ndiye chithunzi cholowetsa
Equation 1.2
Re() imayimira RecCNN. Equation iyi imangopereka mtengo wa equation 1.1 ku RecCNN. ΞΈ amatanthauza magawo ophunzitsidwa bwino a RecCNN (chipewa pamwamba chimatanthauza kuti magawowo akhazikika).
Kutanthauzira mwachilengedwe
Equation 1.0 idzakakamiza ComCNN kusintha zolemera zake kotero kuti, ikamangidwanso pogwiritsa ntchito RecCNN, chithunzi chomaliza chimawoneka chofanana ndi chithunzi cholowetsamo momwe zingathere. Ntchito yachiwiri yotayika ya RecCNN imatanthauzidwa motere:
Equation 2.0
Kufotokozera
Apanso ntchitoyi ikhoza kuwoneka yovuta, koma nthawi zambiri imakhala yokhazikika ya neural network loss function (MSE).
Equation 2.1
Co() amatanthauza kutulutsa kwa codec, x yokhala ndi kapu pamwamba kumatanthauza kutulutsa kwa ComCNN. ΞΈ2 ndi magawo ophunzitsidwa a RecCNN, res() ndikungotulutsa kotsalira kwa RecCNN. Ndizofunikira kudziwa kuti RecCNN imaphunzitsidwa za kusiyana pakati pa Co() ndi chithunzi cholowetsa, koma osati pa chithunzi cholowetsa.
Kutanthauzira mwachilengedwe
Equation 2.0 idzakakamiza RecCNN kuti isinthe zolemera zake kuti zotulutsa ziwoneke ngati zofanana ndi chithunzi cholowera momwe zingathere.
Chiwembu chophunzitsira
Ma Model amaphunzitsidwa mobwerezabwereza, mofanana ndi . Miyezo yachitsanzo choyamba imakonzedwa pamene zolemera zachitsanzo zachiwiri zimasinthidwa, ndiye zolemera zachitsanzo zachiwiri zimakhazikika pamene chitsanzo choyamba chikuphunzitsidwa.
Mayesero
Olembawo anayerekezera njira yawo ndi njira zomwe zilipo, kuphatikizapo ma codec osavuta. Njira yawo imagwira ntchito bwino kuposa ena ndikusunga liwiro lalikulu pa hardware yoyenera. Kuphatikiza apo, olembawo anayesa kugwiritsa ntchito imodzi mwa maukonde awiriwo ndipo adawona kuchepa kwa magwiridwe antchito.
Structural kufanana index (SSIM) poyerekeza. Makhalidwe apamwamba akuwonetsa kufanana kwabwinoko ndi koyambirira. Zotsatira za ntchito ya olemba zimawonetsedwa molimba mtima.
Pomaliza
Tidayang'ana njira yatsopano yogwiritsira ntchito kuphunzira mozama pakuponderezana kwa zithunzi, ndipo tidalankhula za kuthekera kogwiritsa ntchito ma neural network muzochita zopitilira "zambiri", monga kugawa zithunzi ndikusintha zilankhulo. Njirayi sikuti ndi yotsika poyerekeza ndi zofunikira zamakono, komanso imakupatsani mwayi wokonza zithunzi mofulumira kwambiri.
Zakhala zosavuta kuphunzira ma neural network, chifukwa tapanga khodi yotsatsira makamaka kwa okhala ku Khabra HABR, kupereka kuchotsera kwa 10% kuchotsera komwe kwasonyezedwa pachikwangwanicho.
Maphunziro ambiri
Nkhani Zowonetsedwa
Source: www.habr.com