Ama-algorithms aqhutshwa yidatha njengamanethiwekhi e-neural athathe umhlaba wonke. Ukuthuthukiswa kwabo kungenxa yezizathu eziningana, kuhlanganise nemishini eshibhile futhi enamandla kanye nenani elikhulu ledatha. Amanethiwekhi e-Neural okwamanje ahamba phambili kuyo yonke into ehlobene nemisebenzi "yengqondo" efana nokubonwa kwesithombe, ukuqonda kolimi lwemvelo, njll. Kodwa akufanele bagcine emisebenzini enjalo. Lesi sihloko sikhuluma ngendlela yokucindezela izithombe usebenzisa amanethiwekhi e-neural, kusetshenziswa ukufunda okuyinsalela. Indlela evezwe esihlokweni iyashesha futhi ingcono kunamakhodekhi ajwayelekile. Amasu, ama-equations futhi, yiqiniso, itafula elinokuhlolwa ngaphansi kokusikwa.
Lesi sihloko sisekelwe ku umsebenzi. Kucatshangwa ukuthi ujwayelene namanethiwekhi e-neural kanye nemibono yawo. convolution ΠΈ umsebenzi wokulahlekelwa.
Kuyini ukucindezelwa kwesithombe futhi kusebenza kanjani?
Ukuminyanisa isithombe kuyinqubo yokuguqula isithombe ukuze sithathe isikhala esincane. Ukugcina izithombe kalula kuzothatha indawo enkulu, yingakho kunama-codec afana ne-JPEG ne-PNG ahlose ukunciphisa usayizi wesithombe sokuqala.
Njengoba wazi, kunezinhlobo ezimbili zokucindezelwa kwesithombe: akukho ukulahlekelwa ΠΈ ngokulahlekelwa. Njengoba amagama ephakamisa, ukucindezela okungalahleki kungagcina idatha yesithombe sokuqala, kuyilapho ukucindezela okulahlekile kulahlekelwa idatha ethile phakathi nokucindezelwa. isibonelo, i-JPG ama-algorithms alahlekile [approx. transl. - ngokuyisisekelo, masingakhohlwa nge-JPEG engalahleki], futhi i-PNG iyi-algorithm engalahleki.
Ukuqhathaniswa kokucindezelwa okungalahleki nokulahlekelwayo
Qaphela ukuthi maningi ama-artifact angama-blocky esithombeni esingakwesokudla. Lolu ulwazi olulahlekile. Amaphikseli angomakhelwane emibala efanayo acindezelwa njengendawo eyodwa ukonga isikhala, kodwa ulwazi olumayelana namaphikseli angempela lulahlekile. Yiqiniso, ama-algorithms asetshenziswe ku-JPEG, PNG, njll. ama-codec ayinkimbinkimbi kakhulu, kodwa lesi isibonelo esihle esinembile sokucindezelwa kokulahlekelwa. Ukucindezelwa okungenakulahlekelwa kuhle, kodwa amafayela acindezelwe angalahleki athatha isikhala sediski esiningi. Kunezindlela ezingcono zokucindezela izithombe ngaphandle kokulahlekelwa ulwazi oluningi, kodwa zihamba kancane futhi eziningi zisebenzisa izindlela zokuphindaphinda. Lokhu kusho ukuthi azikwazi ukusetshenziswa ngokuhambisana kuma-CPU amaningi noma ama-GPU cores. Lo mkhawulo ubenza bangasebenzi ngokuphelele ekusetshenzisweni kwansuku zonke.
Okokufaka kwe-Convolutional Neural Network
Uma okuthile kudinga ukubalwa futhi izibalo zingalinganiselwa, engeza . Ababhali basebenzise inethiwekhi ye-neural ye-convolutional ejwayelekile ukuze bathuthukise ukucindezelwa kwesithombe. Indlela eyethulwe ayenzi nje kuphela ngokuhambisana nezixazululo ezingcono kakhulu (uma kungenjalo kangcono), ingasebenzisa i-parallel computing, okuholela ekwenyukeni okukhulu kwejubane. Isizathu siwukuthi ama-Convolutional Neural Networks (CNNs) asebenza kahle kakhulu ekukhipheni ulwazi lwendawo ezithombeni, ezibuye zethulwe ngendlela ehlangene (isibonelo, kuphela izingcezu "ezibalulekile" zesithombe ezilondolozwayo). Ababhali bebefuna ukusebenzisa leli khono le-CNN ukumela kangcono izithombe.
bokwakha
Ababhali bahlongoze inethiwekhi ekabili. Inethiwekhi yokuqala ithatha isithombe njengokufakiwe bese ikhiqiza ukumelwa okuhlangene (ComCNN). Okukhiphayo kwale nethiwekhi kube sekucutshungulwa yi-codec ejwayelekile (isb., i-JPEG). Ngemva kokucubungula i-codec, isithombe sidluliselwa kunethiwekhi yesibili, "elungisa" isithombe esivela ku-codec ngomzamo wokubuyisela isithombe sokuqala. Ababhali baqambe le nethiwekhi ngokuthi i-Reconstructive CNN (RecCNN). Njengama-GAN, womabili amanethiwekhi aqeqeshwa ngokuphindaphindiwe.
Ukumelwa kweComCNN Compact kudluliselwa kukhodekhi ejwayelekile
I-RecCNN. Okukhiphayo kwe-ComCNN kuyenyuswa futhi kunikezwe i-RecCNN, ezozama ukufunda okusele
Okukhiphayo kwekhodekhi kuyenyuswa bese kudluliswa ku-RecCNN. I-RecCNN izozama ukwenza isithombe sibe seduze nesangempela ngangokunokwenzeka.
Uhlaka lokucindezelwa kwesithombe oluya ekupheleni. I-Co(.) iyi-algorithm yokuminyanisa isithombe. Ababhali basebenzise i-JPEG, JPEG2000 kanye ne-BPG
Iyini insalela?
Okusele kungacatshangwa njengesinyathelo sangemva kokucubungula "sokuthuthukisa" isithombe esiqoshwa yikhodekhi. Njengoba "inolwazi" oluningi mayelana nomhlaba, inethiwekhi ye-neural ingenza izinqumo ezinengqondo mayelana nokuthi yini okufanele ilungiswe. Lo mbono usekelwe , funda imininingwane ongakwazi ngayo .
Imisebenzi yokulahlekelwa
Imisebenzi emibili yokulahlekelwa isetshenziswa ngoba sinamanethiwekhi amabili e-neural. Eyokuqala yazo, i-ComCNN, ibhalwe ukuthi L1 futhi ichazwa kanje:
Umsebenzi wokulahlekelwa we-ComCNN
Ukuchazwa
Lesi sibalo singabonakala siyinkimbinkimbi, kodwa empeleni siyindinganiso (impande isho iphutha lesikwele) MSE. ||Β² kusho inkambiso yevekhtha abayifakile.
Izibalo 1.1
I-Cr isho ukuphuma kwe-ComCNN. ΞΈ isho ukufundeka kwamapharamitha e-ComCNN, i-XK iyisithombe sokufaka
Izibalo 1.2
Re() imele i-RecCNN. Le zibalo imane idlulisele incazelo yezibalo 1.1 ku-RecCNN. ΞΈ isho amapharamitha e-RecCNN aqeqeshekayo (isigqoko esingaphezulu sisho ukuthi amapharamitha alungisiwe).
Incazelo Enembile
I-Equation 1.0 izobangela i-ComCNN ukuthi iguqule izisindo zayo ukuze kuthi uma iphinda yenziwe nge-RecCNN, isithombe sokugcina sibukeke sifana ngokunokwenzeka nesithombe sokufaka. Umsebenzi wesibili wokulahlekelwa kwe-RecCNN uchazwa kanje:
Izibalo 2.0
Ukuchazwa
Futhi, umsebenzi ungase ubukeke uyinkimbinkimbi, kodwa lokhu ngokuvamile kuwumsebenzi ojwayelekile wokulahleka kwenethiwekhi ye-neural (MSE).
Izibalo 2.1
Co() kusho ukuphuma kwekhodekhi, x ngesigqoko phezulu kusho ukuphuma kwe-ComCNN. ΞΈ2 amapharamitha aqeqeshekayo e-RecCNN, res() iwumphumela nje oyinsalela we-RecCNN. Kuyaphawuleka ukuthi i-RecCNN iqeqeshelwe umehluko phakathi kwe-Co() nesithombe sokufaka, kodwa hhayi esithombeni sokufakwayo.
Incazelo Enembile
I-Equation 2.0 izobangela i-RecCNN ukuthi iguqule izisindo zayo ukuze okukhiphayo kubukeke kufane ngangokunokwenzeka esithombeni okokufaka.
Uhlelo lokufunda
Amamodeli aqeqeshwa ngokuphindaphindiwe, njengokuthi . Izisindo zemodeli yokuqala ziyalungiswa ngenkathi izisindo zemodeli yesibili zibuyekezwa, bese izisindo zemodeli yesibili zilungiswa ngenkathi imodeli yokuqala iqeqeshwa.
Uvivinyo
Ababhali baqhathanise indlela yabo nezindlela ezikhona, ezihlanganisa amakhodekhi alula. Indlela yabo isebenza kangcono kunezinye ngenkathi igcina isivinini esikhulu kuhadiwe ezifanele. Ngaphezu kwalokho, ababhali bazame ukusebenzisa eyodwa kuphela yamanethiwekhi amabili futhi baphawula ukwehla kokusebenza.
Ukuqhathaniswa kwenkomba yokufana kwesakhiwo (i-SSIM). Amanani aphezulu abonisa ukufana okungcono nokwangempela. Uhlobo olugqamile lubonisa umphumela womsebenzi wababhali
isiphetho
Sibheke indlela entsha yokusebenzisa ukufunda okujulile ekucindezelweni kwesithombe, futhi sakhuluma ngethuba lokusebenzisa amanethiwekhi e-neural emisebenzini engaphezu kwemisebenzi βevamileβ efana nokuhlukaniswa kwesithombe nokucubungula ulimi. Le ndlela ayikona nje ukuthi ingaphansi kwezidingo zanamuhla, kodwa futhi ikuvumela ukuthi ucubungule izithombe ngokushesha okukhulu.
Ukufunda amanethiwekhi e-neural sekulula, ngoba senze ikhodi yephromo ikakhulukazi ye-Habravchan I-HABR, enikeza isaphulelo esingeziwe esingu-10% kwisaphulelo esiboniswe kusibhengezo.
Izifundo ezengeziwe
Izihloko Ezifakiwe
Source: www.habr.com