Impendulo emfushane: Ukukhulisa i-AI kusebenza ngokuqeqesha imodeli ezithombeni ezinesinqumo esiphansi nesiphezulu, bese kuyisebenzisa ukubikezela amaphikseli engeziwe angakholeki ngesikhathi sokukhulisa i-AI. Uma imodeli ibone ukuthungwa noma ubuso obufanayo ekuqeqeshweni, ingangeza imininingwane ekholisayo; uma kungenjalo, ingase "ibonise ama-hallucinate" izinto zobuciko ezifana nama-halo, isikhumba esifana ne-wax, noma ukukhanya kuvidiyo.
Izinto ezibalulekile okufanele uzicabangele:
Ukubikezela : Imodeli ikhiqiza imininingwane enengqondo, hhayi ukwakhiwa kabusha okuqinisekisiwe kweqiniso.
Ukukhetha imodeli : Ama-CNN avame ukuba aqinile; ama-GAN angabukeka ebukhali kodwa abe sengozini yokusungula izici.
Ukuhlolwa kwezinto zakudala : Qaphela ama-halo, ukuthungwa okuphindaphindiwe, "cishe izinhlamvu", kanye nobuso obufana nepulasitiki.
Ukuzinza kwevidiyo : Sebenzisa izindlela zesikhashana noma uzobona ukucwazimula nokushukuma kohlaka kuya kuhlaka.
Ukusetshenziswa okunezinzuzo ezinkulu : Uma ukunemba kubalulekile, dalula ukucubungula bese uphatha imiphumela njengesibonelo.
Cishe usuyibonile: isithombe esincane, esiqhekekile siphenduka sibe yinto ecacile ngokwanele ukuphrinta, ukusakaza, noma ukuyifaka esethulweni ngaphandle kokunqekuzisa ikhanda. Kuzwakala njengokukhohlisa. Futhi - ngendlela engcono kakhulu - kufana 😅
Ngakho-ke, indlela i-AI Upscaling esebenza ngayo incike kokuthile okucacile kakhulu kunokuthi “ikhompyutha ithuthukisa imininingwane” (i-hand-wavy) futhi isondele “kumodeli ebikezela isakhiwo esinengqondo sesisombululo esiphezulu ngokusekelwe kumaphethini ewafundile ezibonelweni eziningi” ( Deep Learning for Image Super-resolution: A Survey ). Leso sinyathelo sokubikezela siwumdlalo wonke - futhi yingakho i-AI upscaling ingabukeka imangalisa… noma ipulasitiki encane… noma sengathi ikati lakho likhule ama-bonus whiskers.
Izihloko ongase uthande ukuzifunda ngemva kwalesi:
🔗 Indlela i-AI esebenza ngayo
Funda izisekelo zamamodeli, idatha, kanye nokuphetha ku-AI.
🔗 Indlela i-AI efunda ngayo
Bona ukuthi idatha yokuqeqesha kanye nempendulo kuthuthukisa kanjani ukusebenza kwemodeli ngokuhamba kwesikhathi.
🔗 Indlela i-AI ethola ngayo izinto ezingavamile
Qonda izisekelo zamaphethini nokuthi i-AI ikhombisa kanjani ukuziphatha okungajwayelekile ngokushesha.
🔗 I-AI ibikezela kanjani amathrendi
Hlola izindlela zokubikezela ezibonisa izimpawu futhi zibikezele isidingo sesikhathi esizayo.
Indlela i-AI Upscaling esebenza ngayo: umqondo oyinhloko, ngamazwi avamile 🧩
Ukukhulisa i-scaling kusho ukwandisa isixazululo: amaphikseli amaningi, isithombe esikhulu. Ukukhulisa i-scaling ngokwesiko (njenge-bicubic) ngokuyisisekelo kwelula amaphikseli futhi kusheleleze ukuguquka ( i-Bicubic interpolation ). Kulungile, kodwa ayikwazi ukusungula emisha - imane iyahlanganisa.
Ukukhulisa i-AI kuzama into enesibindi (eyaziwa nangokuthi “isisombululo esiphezulu” ezweni locwaningo) ( Deep Learning for Image Super-resolution: A Survey ):
-
Ibheka okokufaka okune-res ephansi
-
Ubona amaphethini (imiphetho, ukuthungwa, izici zobuso, ukudwetshwa kombhalo, ukuluka kwendwangu…)
-
Kubikezela ukuthi inguqulo ene-res ephezulu kufanele ibukeke kanjani
-
Ikhiqiza idatha eyengeziwe yamaphikseli efanela lawo maphethini
Hhayi "ukubuyisela iqiniso ngokuphelele," njengokuthi "qagela ngendlela engakholeki kakhulu" ( Isithombe Esiphezulu Sokuxazulula Usebenzisa Amanethiwekhi Aqinile E-Convolutional (SRCNN) ). Uma lokho kuzwakala kusolisa kancane, awuphazamanga - kodwa futhi yingakho kusebenza kahle kangaka 😄
Futhi yebo, lokhu kusho ukuthi ukukhushulwa kwe-AI ngokuyisisekelo kuwukulawulwa kokungaboni kahle... kodwa ngendlela ephumelelayo nehlonipha amaphikseli.
Yini eyenza inguqulo enhle yokukhulisa i-AI? ✅🛠️
Uma wahlulela umuntu osebenzisa i-AI esezingeni eliphezulu (noma isethingi esethiwe), nakhu okuvame ukuba yinto ebaluleke kakhulu:
-
Ukulungiswa kwemininingwane ngaphandle kokupheka ngokweqile.
Ukulungiswa kahle kwengeza ukuqina kanye nesakhiwo, hhayi umsindo oqhekekayo noma izikhala mbumbulu. -
Ukuziphatha kahle
kwemigqa ehlanzekile ihlala ihlanzekile. Amamodeli amabi enza imiphetho inyakaze noma ihlume. -
Ukwakheka kwezinwele okungokoqobo
Izinwele akufanele zibe yibhulashi lokupenda. Isitini akufanele sibe yisitembu sephethini esiphindaphindayo. -
Ukuphathwa komsindo nokucindezela
Izithombe eziningi zansuku zonke ziguqulwa zibe yi-JPEG. Umuntu ongcono kakhulu akawukhulisi lowo monakalo ( Real-ESRGAN ). -
Ukuqwashisa ngobuso nombhalo
Ubuso nombhalo yizindawo ezilula kakhulu zokubona amaphutha. Amamodeli amahle awaphatha ngobumnene (noma anezindlela ezikhethekile). -
Ukuvumelana kuzo zonke izinhlaka (zevidiyo)
Uma imininingwane ikhanya kusukela kufreyimu kuya kufreyimu, amehlo akho azokhala. Ividiyo ikhuphula noma ifa ngenxa yokuzinza kwesikhashana ( BasicVSR (CVPR 2021) ). -
Izilawuli ezinengqondo
Ufuna amaslayida ahambisana nemiphumela yangempela: i-denoise, i-deblur, ukususwa kwezinto ezidaliwe, ukugcinwa kokusanhlamvu, ukulola... izinto ezisebenzayo.
Umthetho othule ohlala njalo: ukukhushulwa kwesilinganiso "okungcono kakhulu" kuvame ukuba yilokho ongakuqapheli. Kubukeka sengathi ubunekhamera engcono kakhulu ekuqaleni 📷✨
Ithebula Lokuqhathanisa: izinketho zokukhulisa i-AI ezidumile (nokuthi zilungele ini) 📊🙂
Ngezansi ukuqhathanisa okusebenzayo. Amanani awacacile ngamabomu ngoba amathuluzi ayahlukahluka ngokwelayisensi, amaphakheji, izindleko zokubala, nazo zonke lezo zinto ezijabulisayo.
| Ithuluzi / Indlela | Kuhle kakhulu | Isimo sentengo | Kungani kusebenza (cishe) |
|---|---|---|---|
| Ama-desktop upscale esitayela se-topaz ( Isithombe se-Topaz , Ividiyo ye-Topaz ) | Izithombe, ividiyo, ukuhamba komsebenzi okulula | Ikhokhelwe | Amamodeli ajwayelekile aqinile + ukulungiswa okuningi, avame "ukusebenza nje" ... ikakhulukazi |
| Izici zohlobo lwe-Adobe “Super Resolution” ( Adobe Enhance > Super Resolution ) | Abathwebuli bezithombe kakade bakuleyo ndawo | Ukubhalisa-y | Ukwakhiwa kabusha kwemininingwane eqinile, ngokuvamile okugcina izinto ngendlela ehlelekile (okungadingi idrama) |
| Izinhlobo ze-Real-ESRGAN / ESRGAN ( Real-ESRGAN , ESRGAN ) | DIY, onjiniyela, imisebenzi eminingi | Mahhala (kodwa kubiza isikhathi) | Kuhle kakhulu ngemininingwane yokuthungwa, kungaba mnandi ebusweni uma ungaqaphile |
| Izindlela zokukhulisa usayizi ezisuselwe ekusakazweni ( SR3 ) | Umsebenzi wokudala, imiphumela eyenziwe ngesitayela | Kuxutshiwe | Ngingadala imininingwane emihle - futhi ngingasungula izinto ezingenangqondo, ngakho-ke... yebo |
| Abadlali abaphezulu begeyimu (isitayela se-DLSS/FSR) ( NVIDIA DLSS , AMD FSR 2 ) | Imidlalo yesikhathi sangempela kanye nokuboniswa | Kuhlanganiswe | Isebenzisa idatha yokunyakaza kanye nezinto ezifundiwe - ukusebenza okubushelelezi ukunqoba 🕹️ |
| Izinsizakalo zokukhulisa amafu | Ukunethezeka, ukuwina okusheshayo | Khokha ngokusetshenziswa ngakunye | Kuyashesha + kuyakhula, kodwa uhweba ngokulawula futhi ngezinye izikhathi ubuqili |
| Abakhiqizi be-AI abagxile kumavidiyo ( i-BasicVSR , i-Topaz Video ) | Izithombe ezindala, i-anime, izingobo zomlando | Ikhokhelwe | Amaqhinga esikhashana okunciphisa ukufiphaza + amamodeli evidiyo akhethekile |
| Ukukhulisa ifoni/igalari “ehlakaniphile” | Ukusetshenziswa okuvamile | Kufakiwe | Amamodeli alula alungiselelwe umkhiqizo ojabulisayo, hhayi ukuphelela (kusasebenza) |
Ukufometha ukuvuma okungajwayelekile: “I-Paid-ish” yenza umsebenzi omningi kulelo thebula. Kodwa uthola umbono 😅
Imfihlo enkulu: amamodeli afunda ukumepha kusukela ku-low-res kuya ku-high-res 🧠➡️🖼️
Enhliziyweni yokwenyuka kwe-AI okuningi kukhona ukusethwa kokufunda okugadiwe ( Image Super-Resolution Using Deep Convolutional Networks (SRCNN) ):
-
Qala ngezithombe ezinesinqumo esiphezulu ("iqiniso")
-
Yehlisa isampula ibe yizinguqulo ezinesisombululo esiphansi ("okufakwayo")
-
Qeqesha imodeli yokwakha kabusha i-high-res yokuqala kusuka ku-low-res
Ngokuhamba kwesikhathi, imodeli ifunda ubudlelwano obufana nalokhu:
-
"Lolu hlobo lokufiphala okuzungeze iso ngokuvamile lungolwama-eyelashes"
-
"Leli qembu lamaphikseli livame ukukhombisa umbhalo we-serif"
-
"Lo mphetho we-gradient ubukeka njengomugqa ophahleni, hhayi umsindo ongahleliwe"
Akukona ukubamba ngekhanda izithombe ezithile (ngomqondo olula), ukufunda isakhiwo sezibalo ( Deep Learning for Image Super-resolution: A Survey ). Cabanga ngakho njengokufunda uhlelo lolimi lwezindwangu kanye nemiphetho. Akuyona uhlelo lolimi lwezinkondlo, njenge… uhlelo lolimi lwe-IKEA olusebenzisa ngesandla 🪑📦 (isingathekiso esingaqondakali, kodwa esiseduze ngokwanele).
Ama-nuts nama-bolts: kwenzekani ngesikhathi sokuqagela (uma ukhuphuka) ⚙️✨
Uma ufaka isithombe kumuntu othuthukisa i-AI, ngokuvamile kuba nomzila onjengalona:
-
Ukucubungula kwangaphambili
-
Guqula isikhala sombala (ngezinye izikhathi)
-
Lungisa amanani ephikseli
-
Thayela isithombe sibe yizicucu uma sikhulu (hlola i-VRAM reality 😭) ( I-Real-ESRGAN repo (izinketho zethayela) )
-
-
Ukukhishwa kwesici
-
Izendlalelo zakuqala zibona imiphetho, amakhona, kanye ne-gradients
-
Izendlalelo ezijulile zibona amaphethini: ukuthungwa, izimo, izingxenye zobuso
-
-
Ukwakhiwa kabusha
-
Imodeli ikhiqiza imephu yesici esinokulungiswa okuphezulu
-
Bese uguqula lokho kube umphumela wangempela wephikseli
-
-
Ngemva kokucutshungulwa
-
Ukulola okukhethwa kukho
-
I-denoise ongayikhetha
-
Ukucindezela okukhethwayo kwezinto zobuciko (ukukhala, ama-halo, ukuvinjelwa)
-
Imininingwane eyodwa ecashile: amathuluzi amaningi asezingeni eliphezulu kuma-tile, bese ehlanganisa imithungo. Amathuluzi amahle afihla imingcele yama-tile. Amathuluzi e-Meh ashiya amamaki egridi angabonakali uma ubhekabheka. Futhi yebo, uzobhekabheka, ngoba abantu bathanda ukuhlola ukungapheleli okuncane ku-300% ukusondeza njengama-gremlins amancane 🧌
Imindeni eyinhloko yamamodeli esetshenziselwa ukwandisa i-AI (nokuthi kungani izizwa ihlukile) 🤖📚
1) Isixazululo esiphezulu esisekelwe ku-CNN (ihhashi elisebenzayo lakudala)
Amanethiwekhi e-neural e-Convolutional asebenza kahle kumaphethini endawo: imiphetho, ukuthungwa, izakhiwo ezincane ( Isithombe Esiphezulu Sokuxazulula Ukusebenzisa Amanethiwekhi E-Convolutional Ajulile (SRCNN) ).
-
Izinzuzo: okusheshayo, okuzinzile, izimanga ezimbalwa
-
Ububi: kungabonakala “kucutshungulwa” kancane uma kucindezelwa kanzima
2) Ukwenyusa izinga okusekelwe ku-GAN (isitayela se-ESRGAN) 🎭
Ama-GAN (amaNethiwekhi Aphikisayo Akhiqizayo) aqeqesha ijeneretha ukukhiqiza izithombe ezine-resolution ephezulu umuntu ohlukanisayo angakwazi ukuzihlukanisa kwezangempela ( amaNethiwekhi Aphikisayo Akhiqizayo ).
-
Izinzuzo: imininingwane ehlaba umxhwele, ukuthungwa okumangalisayo
-
Ububi: ingasungula imininingwane engekho - ngezinye izikhathi ayilungile, ngezinye izikhathi ayimangalisi ( SRGAN , ESRGAN )
I-GAN ingakunika lobo bukhali obufanele ukushaywa umoya. Ingaphinde inikeze umuntu othwebula isithombe sakho ishiya elengeziwe. Ngakho-ke… khetha izimpi zakho 😬
3) Ukwenyuka okusekelwe ekusakazweni (i-wildcard yokudala) 🌫️➡️🖼️
Amamodeli okusabalalisa asusa umsindo isinyathelo ngesinyathelo futhi angaqondiswa ukukhiqiza imininingwane ephezulu ( SR3 ).
-
Izinzuzo: kungaba kuhle kakhulu emininingwaneni enengqondo, ikakhulukazi emsebenzini wokudala
-
Ububi: kungasuka kude nobunikazi/isakhiwo sokuqala uma izilungiselelo zinolaka ( SR3 )
Yilapho "ukukhulisa" kuqala khona ukuhlangana kube "ukucabanga kabusha." Ngezinye izikhathi yilokho kanye okufunayo. Ngezinye izikhathi akunjalo.
4) Ukukhulisa ividiyo ngokuhambisana kwesikhathi 🎞️
Ukukhulisa ividiyo kuvame ukunezela ukuqonda okunengqondo kokunyakaza:
-
Isebenzisa ozimele abangomakhelwane ukuze kuqiniswe imininingwane ( BasicVSR (CVPR 2021) )
-
Izama ukugwema izinto ezikhasa phansi nezikhanyayo
-
Ngokuvamile kuhlanganisa i-super-resolution ne-denoise kanye ne-deinterlacing ( i-Topaz Video )
Uma ukukhushulwa kwesithombe kufana nokubuyisela umdwebo owodwa, ukukhushulwa kwevidiyo kufana nokubuyisela i-flipbook ngaphandle kokushintsha isimo sekhala lomlingiswa ekhasini ngalinye. Oku... kunzima kunalokho okuzwakalayo.
Kungani ukukhushulwa kwe-AI ngezinye izikhathi kubonakala kungamanga (nokuthi ungakubona kanjani) 👀🚩
Ukwenyusa i-AI kwehluleka ngezindlela ezibonakalayo. Uma usufunde amaphethini, uzowabona yonke indawo, njengokuthenga imoto entsha bese uqaphela leyo modeli kuzo zonke izitaladi 😵💫
Izinkulumo ezivamile:
-
Isikhumba se-wax ebusweni (umsindo omningi kakhulu + ukushelela)
-
Ama-halo abukhali ngokweqile azungeze imiphetho (indawo "ejwayelekile" ye-overshoot) ( Ukuhlanganiswa kwe-Bicubic )
-
Ukuthungwa okuphindaphindiwe (izindonga zezitini ziba amaphethini okukopisha nokunamathisela)
-
I-micro-contrast ekhungathekisayo ekhala ngokuthi “i-algorithm”
-
Ukubhala okungalungile lapho izinhlamvu ziba cishe izinhlamvu (uhlobo olubi kakhulu)
-
Ukuzulazula kwemininingwane lapho izici ezincane zishintsha khona kancane, ikakhulukazi emisebenzini yokusabalalisa ( SR3 )
Ingxenye eyinkimbinkimbi: ngezinye izikhathi lezi zinto zobuciko zibukeka “zingcono” uma uzibuka nje. Ubuchopho bakho buthanda ubukhali. Kodwa ngemva kwesikhashana, buzizwa… bungasebenzi.
Isu elihle ukusondeza isithombe bese uhlola ukuthi sibukeka semvelo yini ebangeni elijwayelekile lokubuka. Uma sibukeka kahle kuphela ku-zoom engu-400%, lokho akulona ithuba, lokho kuyindlela yokuzilibazisa 😅
Indlela i-AI Upscaling esebenza ngayo: uhlangothi lokuqeqeshwa, ngaphandle kwekhanda lezibalo 📉🙂
Ukuqeqesha amamodeli anezinqumo eziphezulu ngokuvamile kuhilela:
-
Amasethi edatha abhangqiwe (okufakwayo okune-res ephansi, ithagethi ene-res ephezulu) ( Isithombe Esine-Super-Resolution Sisebenzisa Amanethiwekhi E-Deep Convolutional (SRCNN) )
-
Imisebenzi yokulahlekelwa ejezisa ukwakhiwa kabusha okungalungile ( SRGAN )
Izinhlobo zokulahlekelwa ezijwayelekile:
-
Ukulahleka kwe-Pixel (L1/L2)
Kukhuthaza ukunemba. Kungaveza imiphumela ethambile kancane. -
Ukulahlekelwa ukuqonda
Kuqhathanisa izici ezijulile (njengokuthi “ingabe lokhu kubukeka kufana”) kunokuba kuqhathaniswe amaphikseli aqondile ( Ukulahlekelwa ukuqonda (Johnson et al., 2016) ). -
Ukulahlekelwa okuphambene (GAN)
Kukhuthaza ubuqiniso, ngezinye izikhathi ngezindleko zokunemba okungokoqobo ( SRGAN , Generative Adversarial Networks ).
Kukhona ukudonsisana okuqhubekayo:
-
Kwenze kuthembeke kokwasekuqaleni
vs. -
Yenza kube kuhle emehlweni
Amathuluzi ahlukene afika ezindaweni ezahlukene kulolo hlobo. Futhi ungase ukhethe elilodwa kuye ngokuthi ubuyisela izithombe zomndeni noma ulungiselela iphosta lapho "ukubukeka okuhle" kubaluleke kakhulu kunokunemba kwe-forensic.
Imisebenzi esebenzayo: izithombe, ukuskena okudala, i-anime, namavidiyo 📸🧾🎥
Izithombe (izithombe, izindawo, izithombe zomkhiqizo)
Umkhuba omuhle kakhulu uvame ukuba:
-
Umsindo omncane kuqala (uma kudingeka)
-
Okuthuthukisiwe ngezilungiselelo ezilondolozayo
-
Engeza okusanhlamvu uma izinto zizwakala zibushelelezi kakhulu (yebo, ngempela)
Okusanhlamvu kufana nosawoti. Ukudla kwakusihlwa okuningi kakhulu, kodwa akukho nhlobo okunambitheka kancane 🍟
Ukuskena okudala nezithombe ezicindezelwe kakhulu
Lokhu kunzima ngoba imodeli ingase iphathe amabhlogo okucindezela “njengokuthungwa.”
Zama:
-
Ukususwa noma ukususwa kokuvimba kwezinto ezidaliwe
-
Bese kuba ngcono
-
Bese kuba ukulola ukukhanya (hhayi kakhulu... Ngiyazi, wonke umuntu uyakusho lokho, kodwa noma kunjalo)
Ubuciko be-Anime kanye nomugqa
Ubuciko bomugqa buzuza ku:
-
Amamodeli agcina imiphetho ihlanzekile
-
Ukuncipha kokuthungwa kwezinto ezingabonakali
Ukukhushulwa kwe-Anime kuvame ukubukeka kukuhle ngoba izimo zilula futhi zihambisana. (Ngenhlanhla.)
Ividiyo
Ividiyo ingeza izinyathelo ezengeziwe:
-
I-Denoise
-
I-Deinterlace (yemithombo ethile)
-
Okuphezulu
-
Ukushelela noma ukuzinza kwesikhashana ( BasicVSR (CVPR 2021) )
-
Ukufakwa kabusha kokusanhlamvu okukhethwa kukho ukuze kuhlanganiswe
Uma weqa ukuhambisana kwesikhathi, uthola imininingwane ekhazimulayo ikhanya. Uma usuyibonile, awukwazi ukuyisusa. Njengesihlalo esikhalayo ekamelweni elithule 😖
Ukukhetha izilungiselelo ngaphandle kokuqagela ngokunganaki (ishidi elincane lokukhohlisa) 🎛️😵💫
Nasi isimo sengqondo esihle sokuqala:
-
Uma ubuso bubukeka bucwebezela
, nciphisa umsindo, zama imodeli noma imodi yokugcina ubuso. -
Uma ukuthungwa kubonakala kukhulu kakhulu.
Faka amaslayida “okuthuthukisa imininingwane” noma “okubuyisela imininingwane”, ngemva kwalokho engeza okusanhlamvu okuncane. -
Uma imiphetho ikhanya,
nciphisa ukulola phansi, hlola izinketho zokucindezela i-halo. -
Uma isithombe sibukeka “njenge-AI”
Qhubeka ulondoloza kakhulu. Ngezinye izikhathi isinyathelo esihle kakhulu simane nje… sincane.
Futhi: ungasebenzisi i-8x ephezulu ngoba nje ungakwazi. I-2x noma i-4x ehlanzekile ivame ukuba yindawo emnandi. Ngaphambi kwalokho, ucela imodeli ukuthi ibhale inganekwane ngamaphikseli akho 📖😂
Ukuziphatha, ubuqiniso, kanye nombuzo oxakile "weqiniso" 🧭😬
Ukwenyusa i-AI kuqeda umugqa:
-
Ukubuyiselwa kusho ukubuyisa lokho okwakukhona
-
Ukuthuthukisa kusho ukwengeza okwakungenjalo
Ngezithombe zomuntu siqu, ngokuvamile kulungile (futhi kuyathandeka). Ngobuntatheli, ubufakazi bezomthetho, izithombe zezokwelapha, noma yini lapho ukwethembeka kubalulekile khona... udinga ukuqaphela ( OSAC/NIST: Umhlahlandlela Ojwayelekile Wokuphathwa Kwezithombe Zedijithali Ze-Forensic , Iziqondiso ze-SWGDE Zokuhlaziywa Kwezithombe Ze-Forensic ).
Umthetho olula:
-
Uma izinkinga ziphezulu, phatha ukukhushulwa kwe-AI njengokubonisa , hhayi okuqondile.
Futhi, ukudalulwa kubalulekile ezimweni zobungcweti. Hhayi ngoba i-AI imbi, kodwa ngoba izethameli zifanelwe ukwazi ukuthi imininingwane yakhiwe kabusha noma yathathwa. Lokho nje... kuyinhlonipho.
Amanothi okuvala kanye nesifinyezo esifushane 🧡✅
Ngakho-ke, indlela i-AI Upscaling esebenza ngayo yile: amamodeli afunda ukuthi imininingwane enesisombululo esiphezulu ivame ukuhlobana kanjani namaphethini anesisombululo esiphansi, bese ebikezela amaphikseli engeziwe angakholeki ngesikhathi sokukhulisa ( Deep Learning for Image Super-resolution: A Survey ). Kuye ngomndeni wamamodeli (i-CNN, i-GAN, ukusabalala, i-video-temporal), leso sibikezelo singaba esigcinayo nesithembekile… noma sibe nesibindi futhi ngezinye izikhathi singabi nazihibe 😅
Isifinyezo esisheshayo
-
Amaphikseli okukhulisa usayizi wendabuko ( i-Bicubic interpolation )
-
Ukwenyusa i-AI kubikezela imininingwane engekho kusetshenziswa amaphethini afundiwe ( Isithombe Esiphezulu Sokuxazulula Ukusebenzisa Amanethiwekhi Ajulile E-Convolutional (SRCNN) )
-
Imiphumela emihle ivela kumodeli efanele + ukuzibamba
-
Qaphela ama-halo, ubuso obufana ne-wax, ukuthungwa okuphindaphindiwe, kanye nokukhazimula kuvidiyo ( BasicVSR (CVPR 2021) )
-
Ukukhulisa i-scaling ngokuvamile "kuyisakhiwo esisha esingenzeka," hhayi iqiniso eliphelele ( SRGAN , ESRGAN )
Uma ufuna, ngitshele ukuthi yini oyikhulisayo (ubuso, izithombe ezindala, ividiyo, i-anime, ukuskena umbhalo), futhi ngizophakamisa isu lezilungiselelo elivame ukugwema izingibe ezivamile "zokubukeka kwe-AI" 🎯🙂
Imibuzo Evame Ukubuzwa
Ukwenyusa i-AI kanye nendlela esebenza ngayo
Ukwenyusa i-AI (ngokuvamile okubizwa ngokuthi “i-super-resolution”) kwandisa i-resolution yesithombe ngokubikezela imininingwane engekho ye-high resolution evela kumaphethini afundwe ngesikhathi sokuqeqeshwa. Esikhundleni sokumane welule amaphikseli njengokuhlanganiswa kwe-bicubic, imodeli ifunda imiphetho, ukuthungwa, ubuso, kanye nokushaya okufana nombhalo, bese ikhiqiza idatha entsha yephikseli ehambisana nalawo maphethini afundiwe. Akuyona “i-restoration yangempela” kodwa “yenza ukuqagela okukholekayo” okufundeka njengokungokwemvelo.
Ukwenyusa i-AI uma kuqhathaniswa nokushintsha usayizi we-bicubic noma wendabuko
Izindlela zendabuko zokukhulisa izilinganiso (njenge-bicubic) zixubana kakhulu phakathi kwamaphikseli akhona, zisheleleze izinguquko ngaphandle kokudala imininingwane emisha yangempela. Ukukhulisa izilinganiso ze-AI kuhlose ukwakha kabusha isakhiwo esinokwenzeka ngokubona izinkomba ezibonakalayo nokubikezela ukuthi izinguqulo ezisezingeni eliphezulu zalezo zinkomba zivame ukubukeka kanjani. Yingakho imiphumela ye-AI ingazwakala ibukhali kakhulu, kanye nokuthi kungani ingaletha izinto zobuciko noma "isungula" imininingwane ebingekho emthonjeni.
Kungani ubuso bungabonakala bunjengenhlaka noma bubushelelezi ngokweqile
Ubuso obune-wax buvame ukuvela ekususeni umsindo okunamandla kanye nokushelela okuhlanganiswe nokulola okukhumula ukuthungwa kwesikhumba kwemvelo. Amathuluzi amaningi aphatha umsindo kanye nokuthungwa okuhle ngendlela efanayo, ngakho-ke "ukuhlanza" isithombe kungasusa ama-pores kanye nemininingwane ecashile. Indlela evamile ukunciphisa ukuthungwa nokulola, sebenzisa imodi yokugcina ubuso uma ikhona, bese uphinda ulethe ucezu oluthile ukuze umphumela uzwakale ungenapulasitiki futhi uthathwe izithombe.
Izinto zobuciko ezivamile zokukhulisa i-AI okufanele uziqaphele
Ama-tell ajwayelekile afaka phakathi ama-halo azungeze imiphetho, amaphethini okuthungwa okuphindaphindiwe (njengezitini zokukopisha nokunamathisela), i-micro-contrast eqhekeziwe, kanye nombhalo ophenduka "cishe izinhlamvu." Emisebenzini yokusebenza esekelwe ekusakazweni, ungabona futhi ukuzulazula kwemininingwane lapho izici ezincane zishintsha khona kancane. Kuvidiyo, imininingwane ekhanyayo nekhasa kuwo wonke amafreyimu iyizimpawu ezinkulu ezibomvu. Uma ibukeka kahle kuphela ekuzomeni okukhulu, izilungiselelo cishe zinolaka kakhulu.
Indlela i-GAN, i-CNN, kanye nabasabalalisi abasafufusa abavame ukuhluka ngayo ngemiphumela
Isixazululo esiphezulu esisekelwe ku-CNN sivame ukuba sizinzile futhi sibikezeleke kakhudlwana, kodwa singabonakala “sicutshungulwa” uma sicindezelwa kanzima. Izinketho ezisekelwe ku-GAN (isitayela se-ESRGAN) zivame ukukhiqiza ukuthungwa okubukhali nokubonakala okubukhali, kodwa zingaveza imininingwane engalungile, ikakhulukazi ebusweni. Ukwenyuka okusekelwe ekusakazweni kungakhiqiza imininingwane emihle, enengqondo, kodwa kungase kusuke esakhiweni sokuqala uma izilungiselelo zokuqondisa noma zamandla ziqinile kakhulu.
Isu lezilungiselelo ezisebenzayo zokugwema ukubukeka "kwe-AI kakhulu"
Qala ngokulinganisela: 2× noma 4× ephezulu ngaphambi kokufinyelela ezintweni ezibucayi. Uma ubuso bubukeka bucwebezelayo, shayela emuva u-denoise bese ulola bese uzama imodi yokuqaphela ubuso. Uma ukwakheka kuba kukhulu kakhulu, yehlisa ukuthuthukiswa kwemininingwane bese ucabanga ngokungeza okusanhlamvu okuncane ngemva kwalokho. Uma imiphetho ikhanya, yehlisa ukulola bese uhlola i-halo noma ukucindezelwa kwezinto ezidaliwe. Emipayipini eminingi, "okuncane" kuyawina ngoba kulondoloza ubuqiniso obukholekayo.
Ukusingatha amaskeni amadala noma izithombe ezicindezelwe kakhulu nge-JPEG ngaphambi kokukhulisa ubukhulu
Izithombe ezicindezelwe ziyinkimbinkimbi ngoba amamodeli angaphatha izinto zobuciko zamabhulokhi njengokuthungwa kwangempela futhi azikhulise. Indlela evamile yokusebenza ukususa noma ukususa ukuvimba izinto zobuciko kuqala, bese kukhulisa ubukhulu, bese kuba ukulola ukukhanya kuphela uma kudingeka. Kuma-scan, ukuhlanza ngobumnene kungasiza imodeli ukuthi igxile esakhiweni sangempela kunomonakalo. Umgomo ukunciphisa "izikhombisi zokuthungwa mbumbulu" ukuze umuntu ophakeme angaphoqelekile ukwenza ukuqagela okuqinisekile okuvela kokufakwayo okunomsindo.
Kungani ukukhulisa ividiyo kunzima kunokukhulisa isithombe
Ukwenyusa izinga levidiyo kufanele kuhambisane kuzo zonke izinhlaka, hhayi nje kuphela esithombeni esisodwa esimile. Uma imininingwane ishintshashintsha kusuka kufreyimu iye kufreyimu, umphumela uba ophazamisayo ngokushesha. Izindlela ezigxile kuvidiyo zisebenzisa ulwazi lwesikhathi oluvela kuzinhlaka eziseduze ukuze kuqiniswe ukwakhiwa kabusha futhi kugwenywe izinto ezikhazimulayo. Izindlela eziningi zokusebenza zifaka phakathi i-denoise, i-deinterlacing yemithombo ethile, kanye nokufakwa kabusha kokusanhlamvu okukhethwayo ukuze lonke uchungechunge luzwakale luhlangene kunokuba lubukhali ngokwenziwa.
Uma ukukhushulwa kwe-AI kungafaneleki noma kuyingozi ukuthembela kukho
Ukwenyuswa kwe-AI kungcono kakhulu kubhekwa njengokuthuthukiswa, hhayi ubufakazi. Ezimweni ezibaluleke kakhulu njengobuntatheli, ubufakazi bezomthetho, izithombe zezokwelapha, noma umsebenzi we-forensic, ukukhiqiza amaphikseli "akholekayo" kungadukisa ngoba kungangeza imininingwane engathathwanga. Uhlaka oluphephile ukuwusebenzisa ngendlela efanekisayo bese wembula ukuthi inqubo ye-AI yakha kabusha imininingwane. Uma ukuthembeka kubalulekile, gcina okwangempela bese ubhala phansi zonke izinyathelo zokucubungula kanye nezilungiselelo.
Izinkomba
-
arXiv - Ukufunda Okujulile Kwesithombe Isixazululo Esiphezulu: Ucwaningo - arxiv.org
-
arXiv - Isixazululo Esiphezulu Sesithombe Sisebenzisa Amanethiwekhi Ajulile E-Convolutional (SRCNN) - arxiv.org
-
arXiv - Real-ESRGAN - arxiv.org
-
arXiv - ESRGAN - arxiv.org
-
arXiv - SR3 - arxiv.org
-
Unjiniyela we-NVIDIA - I-NVIDIA DLSS - developer.nvidia.com
-
I-AMD GPUOpen - Isixazululo Esiphezulu Se-FidelityFX 2 - gpuopen.com
-
I-Computer Vision Foundation (CVF) Open Access - I-BasicVSR: Ukusesha Izingxenye Ezibalulekile ku-Video Super-Resolution (CVPR 2021) - openaccess.thecvf.com
-
i-arXiv - Amanethiwekhi Aphikisayo Akhiqizayo - arxiv.org
-
arXiv - SRGAN - arxiv.org
-
arXiv - Ukulahlekelwa Kokuqonda (Johnson et al., 2016) - arxiv.org
-
I-GitHub - I-Real-ESRGAN repo (izinketho zethayela) - github.com
-
I-Wikipedia - Ukuhumusha kwe-Bicubic - wikipedia.org
-
Ama-Topaz Labs - Isithombe se-Topaz - topazlabs.com
-
Ama-Topaz Labs - Ividiyo ye-Topaz - topazlabs.com
-
Isikhungo Sosizo se-Adobe - I-Adobe Enhance > Isixazululo Esikhulu - helpx.adobe.com
-
I-NIST / OSAC - Umhlahlandlela Ojwayelekile Wokuphathwa Kwezithombe Zedijithali Ze-Forensic (Inguqulo 1.0) - nist.gov
-
I-SWGDE - Iziqondiso Zokuhlaziywa Kwezithombe Ze-Forensic - swgde.org