Cov ntaub ntawv-tsav algorithms zoo li neural networks tau coj lub ntiaj teb los ntawm cua daj cua dub. Lawv txoj kev loj hlob yog vim muaj ntau yam, suav nrog cov khoom pheej yig thiab muaj zog thiab cov ntaub ntawv loj heev. Neural tes hauj lwm tam sim no nyob rau hauv pem hauv ntej ntawm txhua yam hais txog "kev txawj ntse" cov haujlwm xws li kev paub txog cov duab, kev nkag siab ntawm cov lus ntuj, thiab lwm yam. Tab sis lawv yuav tsum tsis txhob txwv rau cov haujlwm zoo li no. Kab lus no tham txog yuav ua li cas rau compress cov duab siv neural networks, siv cov kev kawm seem. Txoj hauv kev uas tau nthuav tawm hauv kab lus yog nrawm dua thiab zoo dua li tus qauv codecs. Schemes, equations thiab, ntawm chav kawm, ib lub rooj nrog kev ntsuam xyuas nyob rau hauv lub txiav.
Kab lus no yog raws li ua haujlwm. Nws yog xav tias koj paub nrog neural networks thiab lawv cov ntsiab lus. kev sib cav ΠΈ poob haujlwm.
Cov duab compression yog dab tsi thiab nws ua haujlwm li cas?
Duab compression yog cov txheej txheem ntawm kev hloov cov duab kom nws siv qhov chaw tsawg dua. Cias khaws cov duab yuav siv ntau qhov chaw, yog vim li cas thiaj muaj codecs xws li JPEG thiab PNG uas lub hom phiaj txo qhov loj ntawm cov duab qub.
Raws li koj paub, muaj ob hom duab compression: tsis poob ΠΈ nrog poob. Raws li cov npe qhia, lossless compression tuaj yeem khaws cov duab qub cov ntaub ntawv, thaum lossy compression poob qee cov ntaub ntawv thaum compression. Piv txwv li, JPG yog lossy algorithms [approx. txhais. - Hauv paus, cia peb tsis txhob hnov ββββqab txog lossless JPEG], thiab PNG yog qhov tsis muaj kev cuam tshuam.
Kev sib piv ntawm lossless thiab lossy compression
Daim ntawv ceeb toom tias muaj ntau qhov blocky artifacts hauv daim duab ntawm sab xis. Qhov no yog ploj ntaub ntawv. Cov pixels nyob sib ze ntawm cov xim zoo sib xws yog compressed li ib cheeb tsam kom txuag tau qhov chaw, tab sis cov ntaub ntawv hais txog qhov tseeb pixels ploj. Tau kawg, cov algorithms siv nyob rau hauv JPEG, PNG, thiab lwm yam codecs yog ntau nyuaj, tab sis qhov no yog ib tug zoo intuitive piv txwv ntawm lossy compression. Lossless compression yog qhov zoo, tab sis lossless compressed cov ntaub ntawv siv ntau qhov chaw disk. Muaj ntau txoj hauv kev zoo dua los ua kom cov duab tsis poob ntau cov ntaub ntawv, tab sis lawv qeeb heev thiab siv ntau txoj hauv kev. Qhov no txhais tau hais tias lawv tsis tuaj yeem khiav ua ke ntawm ntau lub CPU lossis GPU cores. Qhov kev txwv no ua rau lawv ua tsis tiav hauv kev siv txhua hnub.
Convolutional Neural Network tswv yim
Yog hais tias ib yam dab tsi yuav tsum tau muab xam thiab cov kev xam yuav kwv yees, ntxiv . Cov kws sau ntawv tau siv tus qauv convolutional neural network los txhim kho cov duab compression. Cov txheej txheem nthuav tawm tsis yog tsuas yog ua tau zoo ntawm cov kev daws teeb meem zoo tshaj plaws (yog tias tsis zoo), nws kuj tuaj yeem siv cov kev sib txuas sib txuas, uas ua rau muaj kev nrawm nrawm. Qhov laj thawj yog vim li cas Convolutional Neural Networks (CNNs) zoo heev ntawm kev rho tawm cov ntaub ntawv spatial los ntawm cov duab, uas tau nthuav tawm hauv daim ntawv ntau dua (piv txwv li, tsuas yog "tseem ceeb" ntawm daim duab tau khaws cia). Cov kws sau ntawv xav siv lub peev xwm CNN no los sawv cev rau cov duab zoo dua.
architecture
Cov kws sau ntawv tau thov kom muaj ob lub network. Thawj lub network yuav siv cov duab raws li cov tswv yim thiab tsim ib qho kev sib cog lus sawv cev (ComCNN). Cov zis ntawm lub network no ces ua tiav los ntawm tus qauv codec (xws li JPEG). Tom qab ua tiav los ntawm tus codec, cov duab tau dhau mus rau lub network thib ob, uas "kho" cov duab los ntawm codec hauv kev sim rov qab cov duab qub. Cov neeg sau npe hu ua lub network Reconstructive CNN (RecCNN). Zoo li GANs, ob lub network tau kawm rov ua dua.
ComCNN Compact sawv cev raug xa mus rau tus qauv codec
RecCNN. ComCNN cov zis tau nce thiab pub rau RecCNN, uas yuav sim kawm qhov seem
Cov zis codec tau nce thiab tom qab ntawd dhau mus rau RecCNN. RecCNN yuav sim ua cov duab kom ze rau qhov qub li sai tau.
Xaus-rau-kawg duab compression moj khaum. Co() yog ib qho duab compression algorithm. Cov kws sau ntawv siv JPEG, JPEG2000 thiab BPG
Dab tsi yog qhov seem?
Cov seem tuaj yeem xav tias yog cov kauj ruam tom qab ua "txhim kho" cov duab raug txiav tawm los ntawm codec. Muaj ntau "cov ntaub ntawv" txog lub ntiaj teb no, lub neural network tuaj yeem txiav txim siab txog qhov yuav kho li cas. Lub tswv yim no yog nyob ntawm , nyeem cov ntsiab lus uas koj tuaj yeem ua tau .
Poob functions
Ob txoj haujlwm poob yog siv vim peb muaj ob lub neural networks. Thawj ntawm cov no, ComCNN, tau sau npe L1 thiab tau txhais raws li hauv qab no:
Kev poob haujlwm rau ComCNN
Lus piav qhia
Qhov kev sib npaug no yuav zoo li nyuaj, tab sis nws yog tus qauv (hauv paus txhais tau tias qhov yuam kev) MSEM. ||Β² txhais tau hais tias tus qauv ntawm cov vector lawv nyob ze.
Kev sib npaug 1.1
Cr denotes cov zis ntawm ComCNN. ΞΈ qhia txog kev kawm tau ntawm ComCNN tsis, XK yog cov duab nkag
Kev sib npaug 1.2
Re() sawv cev rau RecCNN. Qhov kev sib npaug no tsuas yog qhia lub ntsiab lus ntawm kab zauv 1.1 rau RecCNN. ΞΈ qhia txog RecCNN cov kev qhia tsis tau (lub kaus mom saum toj txhais tau tias qhov tsis raug kho).
Intuitive txhais
Qhov sib npaug 1.0 yuav ua rau ComCNN hloov nws qhov hnyav kom thaum rov tsim dua nrog RecCNN, cov duab kawg zoo li zoo ib yam li cov duab nkag. Qhov thib ob RecCNN poob muaj nuj nqi yog txhais raws li hauv qab no:
Kev sib npaug 2.0
Lus piav qhia
Ib zaug ntxiv, qhov ua haujlwm yuav zoo li nyuaj, tab sis qhov no yog rau feem ntau tus qauv neural network poob haujlwm (MSE).
Kev sib npaug 2.1
Co() txhais tau tias codec tso zis, x nrog lub kaus mom saum toj txhais tau tias ComCNN tso zis. ΞΈ2 yog RecCNN kev qhia tsis tau, res() tsuas yog RecCNN qhov seem tso zis. Nws yog ib qho tsim nyog sau cia tias RecCNN tau kawm txog qhov sib txawv ntawm Co() thiab cov duab nkag, tab sis tsis yog ntawm cov duab nkag.
Intuitive txhais
Qhov sib npaug 2.0 yuav ua rau RecCNN hloov nws qhov hnyav kom cov zis zoo ib yam li qhov ua tau rau cov duab nkag.
Txoj kev kawm
Cov qauv raug cob qhia rov ua dua, zoo li . Qhov hnyav ntawm thawj tus qauv raug kho thaum qhov hnyav ntawm tus qauv thib ob tab tom hloov kho, ces qhov hnyav ntawm tus qauv thib ob raug kho thaum thawj tus qauv raug cob qhia.
Kev sim
Cov kws sau ntawv piv lawv txoj kev nrog rau txoj kev uas twb muaj lawm, suav nrog cov codecs yooj yim. Lawv txoj kev ua tau zoo dua li lwm tus thaum tswj kev kub ceev ntawm cov khoom siv tsim nyog. Tsis tas li ntawd, cov kws sau ntawv tau sim siv ib qho ntawm ob lub network nkaus xwb thiab sau tseg qhov kev poob qis hauv kev ua haujlwm.
Kev sib piv ntawm cov qauv sib piv (SSIM). Cov txiaj ntsig siab qhia qhov zoo ib yam li qhov qub. Bold hom qhia qhov tshwm sim ntawm kev ua haujlwm ntawm cov kws sau ntawv
xaus
Peb tau saib txoj hauv kev tshiab los siv kev kawm tob rau cov duab compression, thiab tham txog qhov ua tau ntawm kev siv neural networks hauv kev ua haujlwm dhau ntawm "kev ua haujlwm" dav dav xws li kev faib cov duab thiab kev ua cov lus. Txoj kev no tsis yog tsuas yog tsis zoo rau cov kev cai niaj hnub, tab sis kuj tso cai rau koj ua cov duab nrawm dua.
Kev kawm neural tes hauj lwm tau yooj yim dua, vim peb tau ua tus lej tshaj tawm tshwj xeeb rau Habravchan HABR, muab 10% luv nqi ntxiv rau qhov luv nqi qhia rau ntawm daim ntawv lo.
Cov chav kawm ntxiv
Cov ntsiab lus tshwj xeeb
Tau qhov twg los: www.hab.com