Data-inofambiswa algorithms senge neural network yatora nyika nedutu. Kubudirira kwavo kunofambiswa nezvikonzero zvakati wandei, zvinosanganisira zvakachipa uye zvine simba Hardware uye yakawanda data. Neural network parizvino iri pamberi pezvese zvine chekuita ne "cognitive" mabasa akadai sekucherechedzwa kwemifananidzo, kunzwisisa mutauro wechisikigo, nezvimwe. Asi havafaniri kuganhurirwa kumabasa akadaro. Ichi chinyorwa chinotsanangura nzira yekumanikidza mifananidzo uchishandisa neural network uchishandisa yakasara kudzidza. Maitiro anoratidzwa muchinyorwa anoshanda nekukurumidza uye zvirinani pane akajairwa macodecs. Zvirongwa, equations uye, hongu, tafura ine miedzo pasi pekucheka.
Ichi chinyorwa chakavakirwa pa shanda. Zvinofungidzirwa kuti unoziva neural network uye pfungwa dzavo convolution ΠΈ kurasikirwa basa.
Chii chinonzi compression yemufananidzo uye inopinda mumhando dzipi?
Kudzvanya kwemufananidzo inzira yekushandura mufananidzo kuti utore nzvimbo shoma. Kungochengeta mifananidzo kunotora nzvimbo yakawanda, saka kune macodecs akaita seJPEG nePNG anovavarira kuderedza saizi yemufananidzo wekutanga.
Sezvaunoziva, kune marudzi maviri ekumanikidza kwemufananidzo: hapana kurasikirwa ΠΈ nekurasikirwa. Sezvinoratidzwa nemazita, kusarasikirwa kwekumanikidza kunogona kudzoreredza iyo yekutanga mufananidzo data, ukuwo kurasikirwa kwekumanikidza kunorasikirwa nedata panguva yekumanikidza. semuenzaniso, JPG ndeye kurasikirwa algorithms [approx. shanduro - chaizvo, ngatirege zvakare kukanganwa nezve isina kurasikirwa JPEG], uye PNG isingarasikirwe algorithm.
Kufananidza kwekurasikirwa uye kurasikirwa kwekumanikidza
Ziva kuti mufananidzo uri kurudyi une akawanda blocky artifacts. Aya mashoko akarasika. Mapikisi epadhuze emavara akafanana anomanikidzwa senzvimbo imwe kuchengetedza nzvimbo, asi ruzivo rwemapikisi chaiwo rwakarasika. Ehe, maalgorithms anoshandiswa muJPEG, PNG, nezvimwewo macodecs akanyanya kuoma, asi iyi ndiyo yakanaka intuitive muenzaniso wekurasikirwa kwekumanikidza. Kutsikirira kusina kurasikirwa kwakanaka, asi zvisingarasikike kudzvanywa mafaera anotora yakawanda dhisiki nzvimbo. Kune dzimwe nzira dzakasimba dzekudzvanya mifananidzo pasina kurasikirwa neruzivo rwakawanda, asi dzinononoka uye dzakawanda dzinoshandisa nzira dzekudzokorora. Izvi zvinoreva kuti hazvigone kumhanyisa mukufanana pane akawanda CPU kana GPU cores. Ichi chinogumira chinoita kuti zvisashande zvachose pakushandiswa kwemazuva ese.
Convolutional Neural Network Input
Kana chimwe chinhu chichida kuverengwa uye maverengero angave ari ega, wedzera . Vanyori vakashandisa yakaringana yakajairwa convolutional neural network kuvandudza kudzvanywa kwemufananidzo. Iyo yakaratidzwa nzira haingoite pamwe chete neakanakisa mhinduro (kana zvisiri nani), inogona zvakare kushandisa parallel komputa, zvichikonzera kuwedzera kunoshamisa kwekukurumidza. Chikonzero ndechekuti convolutional neural networks (CNNs) yakanaka kwazvo pakubvisa ruzivo rwenzvimbo kubva kumifananidzo, iyo inozomiririrwa mune imwe compact fomu (semuenzaniso, chete "akakosha" mabhiti emufananidzo anochengetwa). Vanyori vaida kushandisa chimiro ichi cheCNN kumiririra zvirinani mifananidzo.
akitekicha
Vanyori vakakurudzira network mbiri. Iyo yekutanga network inotora mufananidzo seyekuisa uye inogadzira inomiririra inomiririra (ComCNN). Kubuda kwetiweki iyi kunozogadziriswa necodec yakajairwa (yakadai seJPEG). Kana yangogadziriswa nekodeki, chifananidzo chinotumirwa kune yechipiri network, iyo "inoruramisa" mufananidzo kubva kucodec mukuedza kudzorera mufananidzo wepakutanga. Vanyori vakadaidza iyi network kuti CNN yekuvakazve (RecCNN). Kufanana nemaGAN, ese ma network anodzidziswa zvakapetwa.
ComCNN Compact inomiririra inoendeswa kune yakajairwa codec
RecCNN. Iyo ComCNN inobuda inokwidziridzwa uye inopihwa kuRecCNN, iyo inoedza kudzidza zvakasara
Iyo codec inobuda inokwidziridzwa uye yobva yapihwa kuRecCNN. RecCNN ichaedza kuburitsa chifananidzo chakafanana nechapakutanga sezvinobvira.
Kupera-kusvika-kumagumo mufananidzo compression framework. Co(.) is an image compression algorithm. Vanyori vakashandisa JPEG, JPEG2000 uye BPG
Chii chasara?
Iyo yasara inogona kufungidzirwa seyeshure-kugadzirisa danho re "kunatsiridza" mufananidzo uri kudhindwa nekodeki. Ne "ruzivo" rwakawanda pamusoro penyika, neural network inogona kuita sarudzo dzekuziva nezve izvo zvekugadzirisa. Pfungwa iyi inobva pane , verenga udzame hwaunogona .
Kurasikirwa mabasa
Mabasa maviri ekurasikirwa anoshandiswa nekuti isu tine maviri neural network. Yekutanga yeiyi, ComCNN, yakanyorwa kuti L1 uye inotsanangurwa seizvi:
Kurasika basa reComCNN
Tsananguro
Iyi equation ingaite seyakaoma, asi ndeye mwero (kureva kukanganisa kwakapetwa kaviri) MSE. ||Β² zvinoreva zvakajairwa nevector yavanovharira.
Equation 1.1
Cr inoreva kubuda kweComCNN. ΞΈ inoreva kudzidziswa kweComCNN paramita, XK ndiyo mufananidzo wekuisa
Equation 1.2
Re() inomirira RecCNN. Iyi equation inongopfuura kukosha kweequation 1.1 kuRecCNN. ΞΈ inoreva maparamendi anodzidziswa eRecCNN (chivharo chiri pamusoro chinoreva kuti paramita yakagadziriswa).
Intuitive tsananguro
Equation 1.0 ichamanikidza ComCNN kuti ichinje uremu hwayo zvekuti, kana ichivakwa patsva uchishandisa RecCNN, mufananidzo wekupedzisira unotaridzika wakafanana nemufananidzo wekuisa sezvinobvira. Yechipiri RecCNN kurasikirwa basa rinotsanangurwa seizvi:
Equation 2.0
Tsananguro
Zvekare basa racho rinogona kutaridzika rakaoma, asi iro rinonyanya chikamu chakajairwa neural network kurasikirwa basa (MSE).
Equation 2.1
Co() zvinoreva kubuda kwecodec, x ine chivharo pamusoro zvinoreva ComCNN kubuda. ΞΈ2 ndiwo anodzidziswa maparamita eRecCNN, res() ingori yasara kubuda kweRecCNN. Zvakakosha kucherechedza kuti RecCNN yakadzidziswa pamusiyano uripo pakati peCo () nemufananidzo wekuisa, asi kwete pamufananidzo wekuisa.
Intuitive tsananguro
Equation 2.0 inomanikidza RecCNN kuti ichinje uremu hwayo kuitira kuti inobuda itaridzike yakafanana nemufananidzo wekuisa sezvinobvira.
Kudzidza chirongwa
Mienzaniso inodzidziswa kakawanda, yakafanana ne . Huremu hwemodhiyo yekutanga hunogadziriswa ukuwo uremu hwemodhi yechipiri huchivandudzwa, kozoti uremu hwemodhiyo yechipiri hunogadzikiswa nepo modhi yekutanga ichidzidziswa.
Miedzo
Vanyori vakafananidza nzira yavo nemaitiro aripo, kusanganisira macodecs akareruka. Nzira yavo inoshanda zvirinani pane vamwe vachichengeta kumhanya kwakanyanya pane yakakodzera hardware. Pamusoro pezvo, vanyori vakaedza kushandisa imwe chete yemanetiweki maviri uye vakaona kudonha kwekuita.
Structural kufanana index (SSIM) kuenzanisa. Hunhu hwepamusoro hunoratidza kufanana kuri nani kune yekutanga. Migumisiro yebasa revanyori inoratidzirwa nemavara matema.
mhedziso
Takatarisa nzira itsva yekushandisa kudzidza kwakadzama kwekumanikidza kwemifananidzo, uye takataura nezve mukana wekushandisa neural network mumabasa anopfuura "akawanda", akadai sekuronga kwemifananidzo uye kugadzirisa mutauro. Iyi nzira haisi chete yakaderera kune zvinodiwa zvemazuva ano, asiwo inokubvumira kugadzirisa mifananidzo nekukurumidza.
Zvave nyore kudzidza neural network, nekuti isu takagadzira kodhi yekusimudzira kunyanya kune vagari veKhabra HABR, ichipa imwezve 10% kuderedzwa kune kuderedzwa kwakaratidzwa pabhena.
Mamwe makosi
Featured Articles
Source: www.habr.com