Ii-algorithms eziqhutywa yidatha ezinje ngothungelwano lwe-neural ziye zathatha ihlabathi ngoqhwithela. Uphuhliso lwabo luqhutywa zizizathu ezininzi, kubandakanywa i-hardware engabizi kwaye enamandla kunye nezixa ezikhulu zedatha. Iinethiwekhi ze-Neural okwangoku ziphambili kuyo yonke into enxulumene "nokuqonda" imisebenzi efana nokuqatshelwa komfanekiso, ukuqonda ulwimi lwendalo, njl. Kodwa abafanele baphelele kwimisebenzi enjalo. Esi sixhobo sichaza indlela yokucinezela imifanekiso kusetyenziswa uthungelwano lwe-neural usebenzisa intsalela yokufunda. Indlela echazwe kwinqaku isebenza ngokukhawuleza kwaye ingcono kune-codecs eziqhelekileyo. Izicwangciso, ii-equations kwaye, ngokuqinisekileyo, itafile eneemvavanyo phantsi kokusikwa.
Eli nqaku lisekelwe kwi umsebenzi. Kucingelwa ukuba uqhelene neenethiwekhi ze-neural kunye neengqikelelo zazo convolution ΠΈ umsebenzi welahleko.
Yintoni ukunyanzeliswa komfanekiso kwaye zeziphi iintlobo ezingena kuzo?
Uxinzelelo lomfanekiso yinkqubo yokuguqula umfanekiso ukuze uthabathe indawo encinci. Ukugcina nje imifanekiso kuya kuthatha indawo eninzi, ngoko ke kukho iicodecs ezifana neJPEG kunye nePNG ezijolise ekunciphiseni ubungakanani bomfanekiso wokuqala.
Njengoko usazi, zimbini iindidi zokucinezelwa komfanekiso: akukho lahleko ΠΈ ngeelahleko. Njengoko amagama ecebisa, ucinezelo olungalahlekiyo lunokubuyisela idatha yoqobo yomfanekiso, ngelixa ucinezelo lwelahleko luphulukana nedatha ethile ngexesha loxinzelelo. umzekelo, iJPG zialgorithms ezilahlekileyo [approx. inguqulelo - ngokusisiseko, masingalibali ngeJPEG engalahlekiyo], kwaye i-PNG yi-algorithm engalahlekiyo.
Ukuthelekiswa koxinzelelo olungenakulahlekelwa kunye nelahleko
Qaphela ukuba umfanekiso osekunene unee-artifacts ezininzi ezibhloko. Olu lulwazi olulahlekileyo. Iiphikseli ezikufutshane zemibala efanayo zicinezelwa njengendawo enye yokugcina indawo, kodwa ulwazi malunga neepixels zokwenyani lulahlekile. Ngokuqinisekileyo, i-algorithms esetyenziswe kwiJPEG, i-PNG, njl.njl. i-codecs inzima kakhulu, kodwa lo ngumzekelo omhle we-intuitive wokunyanzeliswa kwelahleko. Uxinzelelo olungenalahleko lulungile, kodwa iifayile ezicinezelekileyo zithatha indawo eninzi yedisk. Kukho iindlela ezisebenzayo zokucinezela imifanekiso ngaphandle kokuphulukana nolwazi oluninzi, kodwa ziyacotha kwaye uninzi zisebenzisa iindlela zokuphindaphinda. Oku kuthetha ukuba abanakuqhutywa ngokunxuseneyo kwii-CPU ezininzi okanye ii-GPU cores. Lo mda ubenza bangasebenzi ngokupheleleyo kusetyenziso lwemihla ngemihla.
Ungeniso lweNeural Neural Network
Ukuba kukho into efuna ukubalwa kwaye izibalo zinokuqikelelwa, yongeza . Ababhali basebenzise inethiwekhi ye-neural ye-convolutional esemgangathweni ukuphucula ukucinezelwa komfanekiso. Indlela enikezelweyo ayisebenzi kuphela ngokuhambelana nezona zisombululo ezilungileyo (ukuba azingcono), inokusebenzisa i-parallel computing, ekhokelela ekunyukeni okukhulu kwesantya. Isizathu sesokuba i-convolutional neural networks (CNNs) zilunge kakhulu ekukhupheni ulwazi lwendawo kwimifanekiso, ethi ke imelwe kwifom edityanisiweyo (umzekelo, kuphela "izinto ezibalulekileyo" zomfanekiso ezigciniweyo). Ababhali bafuna ukusebenzisa eli nqaku le-CNN ukumela ngcono imifanekiso.
izakhiwo
Ababhali bacebise inethiwekhi ezimbini. Inethiwekhi yokuqala ithatha umfanekiso njengegalelo kwaye ivelise umboniso ohlangeneyo (ComCNN). Imveliso yalo msebenzi wothungelwano iphinde iqhutywe yikhowudi eqhelekileyo (efana neJPEG). Emva kokuba iqhutywe yi-codec, umfanekiso uthunyelwa kumnatha wesibini, othi "ulungise" umfanekiso ovela kwi-codec kumzamo wokubuyisela umfanekiso wokuqala. Ababhali babiza le nethiwekhi ngokuba yi-CNN yokwakhiwa kwakhona (RecCNN). Njengee-GAN, zombini iinethiwekhi ziqeqeshwa ngokuphindaphindiweyo.
IComCNN Compact emele ikhutshelwe kwicodec eqhelekileyo
I-RecCNN. Imveliso yeComCNN inyusiwe kwaye yondliwa kwi-RecCNN, eya kuzama ukufunda intsalela
Imveliso ye-codec inyuswa kwaye yondliwa kwi-RecCNN. I-RecCNN iya kuzama ukuvelisa umfanekiso ofana nowokuqala kangangoko.
Isakhelo sokucinezela umfanekiso wokuya ekupheleni. I-Co (.) yi-algorithm yoxinzelelo lomfanekiso. Ababhali basebenzisa iJPEG, JPEG2000 kunye neBPG
Ithini intsalela?
Intsalela inokucingelwa njengenyathelo emva kokulungiswa "kokuphucula" umfanekiso ohlanjululwa yi-codec. Ngoninzi "lolwazi" malunga nehlabathi, inethiwekhi ye-neural inokwenza izigqibo zengqondo malunga nokuba yintoni ukulungisa. Le ngcamango isekelwe , funda iinkcukacha onokuthi ngazo .
Ilahleko imisebenzi
Imisebenzi emibini yelahleko isetyenzisiweyo kuba sineenethiwekhi ezimbini ze-neural. Eyokuqala kwezi, i-ComCNN, ibhalwe L1 kwaye ichazwa ngolu hlobo lulandelayo:
Umsebenzi welahleko weComCNN
Inkcazo
Le nxaki inokubonakala inzima, kodwa isemgangathweni (ithetha impazamo ephindwe kabini) MSE. ||Β² ithetha isiqhelo sevektha abayivaleleyo.
Inxaki 1.1
I-Cr ichaza imveliso yeComCNN. ΞΈ ichaza ukuqeqesheka kweeparamitha ze-ComCNN, i-XK ngumfanekiso wegalelo
Inxaki 1.2
Re() imele i-RecCNN. Le nxaki igqithisa nje ixabiso lenxaki 1.1 kwiRecCNN. ΞΈ ichaza iiparamitha eziqeqeshekayo ze-RecCNN (i-cap phezulu ithetha ukuba iiparamitha zilungisiwe).
Inkcazo eqondakalayo
I-Equation 1.0 iya kunyanzela i-ComCNN ukuba itshintshe iintsimbi zayo ukuze, xa iphinda yakhiwe kusetyenziswa i-RecCNN, umfanekiso wokugqibela ubukeka ufana nomfanekiso wegalelo kangangoko kunokwenzeka. Umsebenzi wesibini weRecCNN welahleko uchazwa ngolu hlobo lulandelayo:
Inxaki 2.0
Inkcazo
Kwakhona umsebenzi unokujongeka untsonkothile, kodwa ubukhulu becala buyilahleko yomsebenzi womnatha we-neural (MSE).
Inxaki 2.1
Co() ithetha imveliso yecodec, x ene-cap phezulu ithetha imveliso yeComCNN. ΞΈ2 ziiparamitha eziqeqeshekayo zeRecCNN, res() yintsalela yesiphumo seRecCNN. Kuyaphawuleka ukuba i-RecCNN iqeqeshwe kumahluko phakathi kwe-Co () kunye nomfanekiso wegalelo, kodwa kungekhona kumfanekiso wegalelo.
Inkcazo eqondakalayo
I-Equation 2.0 iya kunyanzela i-RecCNN ukuba itshintshe ubunzima bayo ukuze imveliso ibonakale ifana nomfanekiso wegalelo kangangoko kunokwenzeka.
Inkqubo yokufunda
Iimodeli ziqeqeshwa ngokuphindaphindiweyo, ngokufanayo . Ubunzima bemodeli yokuqala bulungiswa ngelixa iintsimbi zemodeli yesibini zihlaziywa, emva koko iintsimbi zemodeli yesibini zilungiswa ngelixa imodeli yokuqala iqeqeshwa.
Iimvavanyo
Ababhali bathelekisa indlela yabo kunye neendlela ezikhoyo, kuquka i-codecs elula. Indlela yabo isebenza ngcono kunabanye ngelixa igcina isantya esiphezulu kwi-hardware efanelekileyo. Ukongezelela, ababhali bazama ukusebenzisa enye kuphela yenethiwekhi ezimbini kwaye baqaphele ukuhla ekusebenzeni.
Isalathiso sokufana kwesakhiwo (SSIM) uthelekiso. Amaxabiso aphezulu abonisa ukufana okungcono kweyokuqala. Iziphumo zomsebenzi wababhali ziphawulwe ngokungqindilili.
isiphelo
Sijonge indlela entsha yokusebenzisa ukufunda okunzulu kuxinzelelo lwemifanekiso, kwaye sathetha malunga nokwenzeka kokusebenzisa uthungelwano lwe-neural kwimisebenzi engaphaya kwe "jikelele", efana nokuhlelwa kwemifanekiso kunye nokulungiswa kolwimi. Le ndlela ayikho ngaphantsi kweemfuno zanamhlanje, kodwa ikuvumela ukuba uqhube imifanekiso ngokukhawuleza.
Kuye kube lula ukufunda uthungelwano lwe-neural, kuba senze ikhowudi yokuthengisa ngakumbi kubahlali baseKhabra IHABR, ukunika isaphulelo esongezelelweyo se-10% kwisaphulelo esiboniswe kwibhena.
Izifundo ezingakumbi
Amanqaku akhoyo
umthombo: www.habr.com