G\u00f6r\u00fcnt\u00fc verileri, modern dijital \u00e7a\u011f\u0131n en yayg\u0131n ve hacimli i\u00e7erik t\u00fcrlerinden biridir. Ak\u0131ll\u0131 telefonlar\u0131m\u0131zdan profesyonel kameralara, t\u0131bbi g\u00f6r\u00fcnt\u00fcleme cihazlar\u0131ndan uydu sens\u00f6rlerine kadar her alanda devasa miktarda g\u00f6rsel bilgi \u00fcretilmektedir. Bu verinin depolanmas\u0131, iletilmesi ve i\u015flenmesi, \u00f6nemli bant geni\u015fli\u011fi, depolama alan\u0131 ve i\u015flem g\u00fcc\u00fc gereksinimleri do\u011furmaktad\u0131r. \u0130\u015fte tam da bu noktada g\u00f6r\u00fcnt\u00fc s\u0131k\u0131\u015ft\u0131rma teknolojileri devreye girer. G\u00f6r\u00fcnt\u00fc s\u0131k\u0131\u015ft\u0131rma, g\u00f6rsel bilginin gereksiz fazlal\u0131klar\u0131n\u0131 ortadan kald\u0131rarak veya azaltarak, ayn\u0131 bilginin daha az veri ile temsil edilmesini sa\u011flar. Bu sayede depolama maliyetleri d\u00fc\u015fer, iletim h\u0131zlar\u0131 artar ve genel sistem performans\u0131 iyile\u015fir.<\/p>\n
Geleneksel g\u00f6r\u00fcnt\u00fc s\u0131k\u0131\u015ft\u0131rma y\u00f6ntemleri, JPEG, PNG, GIF gibi standartlarla uzun y\u0131llard\u0131r kullan\u0131lmaktad\u0131r. JPEG (Joint Photographic Experts Group), \u00f6zellikle foto\u011frafik g\u00f6r\u00fcnt\u00fcler i\u00e7in tasarlanm\u0131\u015f kay\u0131pl\u0131 bir s\u0131k\u0131\u015ft\u0131rma format\u0131d\u0131r ve insan g\u00f6z\u00fcn\u00fcn alg\u0131lama e\u015fi\u011fini kullanarak y\u00fcksek s\u0131k\u0131\u015ft\u0131rma oranlar\u0131 sunar. PNG (Portable Network Graphics) ise kay\u0131ps\u0131z s\u0131k\u0131\u015ft\u0131rma i\u00e7in tercih edilir ve \u00f6zellikle grafikler, logolar gibi keskin hatl\u0131 g\u00f6r\u00fcnt\u00fclerde kalite kayb\u0131n\u0131 \u00f6nler. Bu y\u00f6ntemler olduk\u00e7a ba\u015far\u0131l\u0131 olsa da, sabit algoritmalar kullanmalar\u0131 nedeniyle her zaman optimum s\u0131k\u0131\u015ft\u0131rma oran\u0131n\u0131 ve kaliteyi sa\u011flamayabilirler. \u00d6zellikle karma\u015f\u0131k ve dinamik g\u00f6r\u00fcnt\u00fc i\u00e7eriklerinde veya belirli uygulama senaryolar\u0131nda, daha esnek ve veriye \u00f6zg\u00fc s\u0131k\u0131\u015ft\u0131rma yakla\u015f\u0131mlar\u0131na ihtiya\u00e7 duyulmaktad\u0131r.<\/p>\n
Son y\u0131llarda derin \u00f6\u011frenme alan\u0131ndaki h\u0131zl\u0131 geli\u015fmeler, g\u00f6r\u00fcnt\u00fc s\u0131k\u0131\u015ft\u0131rma problemlerine yeni ve heyecan verici \u00e7\u00f6z\u00fcmler sunmu\u015ftur. Derin \u00f6\u011frenme modelleri, b\u00fcy\u00fck veri k\u00fcmelerinden \u00f6\u011frenme yetenekleri sayesinde, geleneksel y\u00f6ntemlerin \u00f6tesine ge\u00e7ebilen, adaptif ve y\u00fcksek performansl\u0131 s\u0131k\u0131\u015ft\u0131rma algoritmalar\u0131 geli\u015ftirmemizi sa\u011flam\u0131\u015ft\u0131r. Bu yakla\u015f\u0131mlardan biri de otoenkoderlerdir. Otoenkoderler, bir girdiyi s\u0131k\u0131\u015ft\u0131r\u0131lm\u0131\u015f bir temsile (gizli uzay veya latent space) d\u00f6n\u00fc\u015ft\u00fcrmek ve ard\u0131ndan bu s\u0131k\u0131\u015ft\u0131r\u0131lm\u0131\u015f temsilden orijinal girdiyi yeniden olu\u015fturmak i\u00e7in tasarlanm\u0131\u015f \u00f6zel bir yapay sinir a\u011f\u0131 t\u00fcr\u00fcd\u00fcr. G\u00f6r\u00fcnt\u00fc s\u0131k\u0131\u015ft\u0131rma ba\u011flam\u0131nda, otoenkoderler, g\u00f6r\u00fcnt\u00fclerin en temel ve anlaml\u0131 \u00f6zelliklerini \u00f6\u011frenerek, bu \u00f6zellikleri d\u00fc\u015f\u00fck boyutlu bir vekt\u00f6r olarak kodlar ve ard\u0131ndan bu vekt\u00f6rden orijinal g\u00f6r\u00fcnt\u00fcy\u00fc m\u00fcmk\u00fcn olan en y\u00fcksek kalitede yeniden in\u015fa etmeye \u00e7al\u0131\u015f\u0131r.<\/p>\n
Bu makalede, Keras derin \u00f6\u011frenme k\u00fct\u00fcphanesini kullanarak otoenkoderlerin g\u00f6r\u00fcnt\u00fc s\u0131k\u0131\u015ft\u0131rma amac\u0131yla nas\u0131l kullan\u0131labilece\u011fini detayl\u0131 bir \u015fekilde inceleyece\u011fiz. Otoenkoderlerin temel prensiplerinden ba\u015flayarak, farkl\u0131 t\u00fcrlerini, evri\u015fimsel otoenkoderlerin g\u00f6r\u00fcnt\u00fc s\u0131k\u0131\u015ft\u0131rmadaki \u00f6nemini, Keras ile bir otoenkoder modelinin nas\u0131l olu\u015fturulaca\u011f\u0131n\u0131, e\u011fitilece\u011fini ve de\u011ferlendirilece\u011fini ad\u0131m ad\u0131m a\u00e7\u0131klayaca\u011f\u0131z. Ayr\u0131ca, bu yakla\u015f\u0131m\u0131n avantajlar\u0131n\u0131 ve dezavantajlar\u0131n\u0131 tart\u0131\u015farak, gelecekteki potansiyel geli\u015fmelere de de\u011finece\u011fiz.<\/p>\n
G\u00f6r\u00fcnt\u00fc s\u0131k\u0131\u015ft\u0131rma, g\u00f6r\u00fcnt\u00fc verisinin boyutunu k\u00fc\u00e7\u00fcltme i\u015flemidir. Bu i\u015flem, genellikle g\u00f6r\u00fcnt\u00fcn\u00fcn i\u00e7erdi\u011fi gereksiz veya tekrarlayan bilgileri ortadan kald\u0131rarak ger\u00e7ekle\u015ftirilir. G\u00f6r\u00fcnt\u00fc s\u0131k\u0131\u015ft\u0131rma algoritmalar\u0131 iki ana kategoriye ayr\u0131l\u0131r:<\/p>\n
Kay\u0131ps\u0131z s\u0131k\u0131\u015ft\u0131rma y\u00f6ntemleri, s\u0131k\u0131\u015ft\u0131r\u0131lm\u0131\u015f g\u00f6r\u00fcnt\u00fcn\u00fcn orijinal g\u00f6r\u00fcnt\u00fc ile tamamen ayn\u0131 olmas\u0131n\u0131 sa\u011flar. Yani, s\u0131k\u0131\u015ft\u0131rma ve ard\u0131ndan a\u00e7ma i\u015flemi sonras\u0131nda hi\u00e7bir bilgi kayb\u0131 ya\u015fanmaz. Bu t\u00fcr s\u0131k\u0131\u015ft\u0131rma, t\u0131bbi g\u00f6r\u00fcnt\u00fcler, teknik \u00e7izimler veya metin i\u00e7eren g\u00f6r\u00fcnt\u00fcler gibi veri b\u00fct\u00fcnl\u00fc\u011f\u00fcn\u00fcn kritik oldu\u011fu alanlarda tercih edilir. PNG, GIF ve TIFF gibi formatlar kay\u0131ps\u0131z s\u0131k\u0131\u015ft\u0131rma kullan\u0131r. Ancak, kay\u0131ps\u0131z s\u0131k\u0131\u015ft\u0131rma oranlar\u0131 genellikle kay\u0131pl\u0131 s\u0131k\u0131\u015ft\u0131rmaya g\u00f6re daha d\u00fc\u015f\u00fckt\u00fcr.<\/p>\n
Kay\u0131pl\u0131 s\u0131k\u0131\u015ft\u0131rma y\u00f6ntemleri, g\u00f6r\u00fcnt\u00fcden insan g\u00f6z\u00fcn\u00fcn alg\u0131lamakta zorland\u0131\u011f\u0131 veya alg\u0131lamad\u0131\u011f\u0131 baz\u0131 bilgileri kal\u0131c\u0131 olarak kald\u0131r\u0131r. Bu sayede \u00e7ok daha y\u00fcksek s\u0131k\u0131\u015ft\u0131rma oranlar\u0131 elde edilebilir. Ancak, s\u0131k\u0131\u015ft\u0131rma ve a\u00e7ma i\u015flemi sonras\u0131nda orijinal g\u00f6r\u00fcnt\u00fcden bir miktar sapma meydana gelir. JPEG, s\u0131k\u0131\u015ft\u0131rma oran\u0131n\u0131 g\u00f6r\u00fcnt\u00fc kalitesiyle dengeleyerek en yayg\u0131n kay\u0131pl\u0131 s\u0131k\u0131\u015ft\u0131rma format\u0131d\u0131r. Otoenkoderler genellikle kay\u0131pl\u0131 s\u0131k\u0131\u015ft\u0131rma prensibiyle \u00e7al\u0131\u015f\u0131r, \u00e7\u00fcnk\u00fc ama\u00e7lar\u0131 verinin en \u00f6nemli \u00f6zelliklerini koruyarak en d\u00fc\u015f\u00fck boyutlu temsili elde etmektir.<\/p>\n
S\u0131k\u0131\u015ft\u0131rma oran\u0131, orijinal dosya boyutunun s\u0131k\u0131\u015ft\u0131r\u0131lm\u0131\u015f dosya boyutuna b\u00f6l\u00fcnmesiyle elde edilir. Y\u00fcksek bir s\u0131k\u0131\u015ft\u0131rma oran\u0131, daha az depolama alan\u0131 ve bant geni\u015fli\u011fi anlam\u0131na gelir. Ancak, kay\u0131pl\u0131 s\u0131k\u0131\u015ft\u0131rmada y\u00fcksek s\u0131k\u0131\u015ft\u0131rma oran\u0131 genellikle daha fazla kalite kayb\u0131na yol a\u00e7ar. Bu nedenle, s\u0131k\u0131\u015ft\u0131rma algoritmalar\u0131n\u0131 de\u011ferlendirirken hem s\u0131k\u0131\u015ft\u0131rma oran\u0131n\u0131 hem de g\u00f6r\u00fcnt\u00fc kalitesini g\u00f6z \u00f6n\u00fcnde bulundurmak \u00f6nemlidir.<\/p>\n
G\u00f6r\u00fcnt\u00fc kalitesini nicel olarak de\u011ferlendirmek i\u00e7in \u00e7e\u015fitli metrikler kullan\u0131l\u0131r: Otoenkoderler, girdilerini s\u0131k\u0131\u015ft\u0131r\u0131lm\u0131\u015f bir temsile d\u00f6n\u00fc\u015ft\u00fcrmek ve bu temsilden girdilerini m\u00fcmk\u00fcn oldu\u011funca asl\u0131na uygun bir \u015fekilde yeniden yap\u0131land\u0131rmak i\u00e7in e\u011fitilmi\u015f bir t\u00fcr yapay sinir a\u011f\u0131d\u0131r. Denetimsiz \u00f6\u011frenme (unsupervised learning) paradigmalar\u0131 alt\u0131nda s\u0131n\u0131fland\u0131r\u0131l\u0131rlar, \u00e7\u00fcnk\u00fc etiketli verilere ihtiya\u00e7 duymazlar; bunun yerine girdinin kendisi hem giri\u015f hem de beklenen \u00e7\u0131k\u0131\u015f olarak kullan\u0131l\u0131r.<\/p>\n Bir otoenkoderin temel mimarisi iki ana b\u00f6l\u00fcmden olu\u015fur: Otoenkoderin amac\u0131, \u00e7\u0131k\u0131\u015f\u0131n giri\u015fle m\u00fcmk\u00fcn oldu\u011funca benzer olmas\u0131n\u0131 sa\u011flamakt\u0131r. Bu “benzerlik” genellikle bir kay\u0131p fonksiyonu (loss function) ile \u00f6l\u00e7\u00fcl\u00fcr. G\u00f6r\u00fcnt\u00fc s\u0131k\u0131\u015ft\u0131rma ba\u011flam\u0131nda, en yayg\u0131n kay\u0131p fonksiyonu Ortalama Kare Hata (MSE) veya \u0130kili \u00c7apraz Entropi (Binary Cross-Entropy) gibi fonksiyonlard\u0131r. Model, bu kay\u0131p fonksiyonunu minimize etmek i\u00e7in e\u011fitilir.<\/p>\n Gizli uzay, otoenkoderin en kritik bile\u015fenidir. Giri\u015f verisinin, yani g\u00f6r\u00fcnt\u00fcn\u00fcn, d\u00fc\u015f\u00fck boyutlu ve yo\u011fun bir temsilidir. Bu uzaydaki her bir nokta, potansiyel bir g\u00f6r\u00fcnt\u00fc veya g\u00f6r\u00fcnt\u00fcn\u00fcn bir \u00f6zelli\u011fini temsil eder. Otoenkoderin \u00f6\u011frenme s\u00fcreci boyunca, a\u011f, gizli uzay\u0131n her bir boyutunun giri\u015f verisindeki hangi varyasyonu temsil etti\u011fini \u00f6\u011frenir. G\u00f6r\u00fcnt\u00fc s\u0131k\u0131\u015ft\u0131rma i\u00e7in, bu gizli uzay\u0131n boyutu, s\u0131k\u0131\u015ft\u0131r\u0131lm\u0131\u015f g\u00f6r\u00fcnt\u00fcn\u00fcn boyutunu do\u011frudan belirler. Gizli uzay\u0131n boyutu ne kadar k\u00fc\u00e7\u00fck olursa, s\u0131k\u0131\u015ft\u0131rma oran\u0131 o kadar y\u00fcksek olur; ancak bu, yeniden olu\u015fturulan g\u00f6r\u00fcnt\u00fcn\u00fcn kalitesinde potansiyel bir d\u00fc\u015f\u00fc\u015fe de yol a\u00e7abilir.<\/p>\n Otoenkoderler, veri s\u0131k\u0131\u015ft\u0131rman\u0131n yan\u0131 s\u0131ra, g\u00fcr\u00fclt\u00fc giderme (denoising), anomali tespiti ve \u00fcretken modelleme gibi bir\u00e7ok farkl\u0131 g\u00f6revde de kullan\u0131labilir. Temel fikir, verinin \u00f6z\u00fcn\u00fc yakalamak ve bu \u00f6z\u00fc kullanarak hem orijinal veriyi yeniden olu\u015fturmak hem de potansiyel olarak yeni veriler \u00fcretmektir.<\/p>\n Otoenkoderler, temel mimarinin \u00f6tesinde, belirli g\u00f6revler veya veri t\u00fcrleri i\u00e7in optimize edilmi\u015f bir\u00e7ok farkl\u0131 varyasyona sahiptir. G\u00f6r\u00fcnt\u00fc s\u0131k\u0131\u015ft\u0131rma alan\u0131nda, \u00f6zellikle evri\u015fimsel otoenkoderler b\u00fcy\u00fck \u00f6nem ta\u015f\u0131r.<\/p>\n En temel otoenkoderler, tamamen ba\u011fl\u0131 (fully connected) katmanlardan olu\u015fur. Kodlay\u0131c\u0131, giri\u015f g\u00f6r\u00fcnt\u00fcs\u00fcn\u00fc d\u00fczle\u015ftirilmi\u015f bir vekt\u00f6r olarak al\u0131r ve bir dizi yo\u011fun katman arac\u0131l\u0131\u011f\u0131yla gizli uzaya indirger. Kod \u00e7\u00f6z\u00fcc\u00fc ise gizli uzay vekt\u00f6r\u00fcn\u00fc tekrar d\u00fczle\u015ftirilmi\u015f bir g\u00f6r\u00fcnt\u00fc vekt\u00f6r\u00fcne d\u00f6n\u00fc\u015ft\u00fcr\u00fcr ve ard\u0131ndan bu vekt\u00f6r g\u00f6r\u00fcnt\u00fc boyutlar\u0131na yeniden \u015fekillendirilir. Bu t\u00fcr otoenkoderler, MNIST gibi k\u00fc\u00e7\u00fck, basit g\u00f6r\u00fcnt\u00fcler i\u00e7in \u00e7al\u0131\u015fabilirken, daha b\u00fcy\u00fck ve karma\u015f\u0131k g\u00f6r\u00fcnt\u00fclerde uzamsal bilgiyi korumakta yetersiz kal\u0131r ve \u00e7ok say\u0131da parametreye ihtiya\u00e7 duyar. Bu nedenle, g\u00f6r\u00fcnt\u00fc s\u0131k\u0131\u015ft\u0131rma i\u00e7in genellikle tercih edilmezler.<\/p>\n Evri\u015fimsel otoenkoderler, g\u00f6r\u00fcnt\u00fc s\u0131k\u0131\u015ft\u0131rma i\u00e7in en uygun otoenkoder t\u00fcr\u00fcd\u00fcr. Geleneksel evri\u015fimli sinir a\u011flar\u0131n\u0131n (CNN) \u00f6zelliklerini kullanarak, g\u00f6r\u00fcnt\u00fclerin uzamsal hiyerar\u015fisini ve yerel \u00f6zelliklerini daha etkili bir \u015fekilde \u00f6\u011frenirler.<\/p>\n * Kodlay\u0131c\u0131:<\/strong> Evri\u015fimsel kodlay\u0131c\u0131, bir dizi CAE’ler, g\u00f6r\u00fcnt\u00fc verileriyle \u00e7al\u0131\u015f\u0131rken basit otoenkoderlere g\u00f6re \u00e7ok daha ba\u015far\u0131l\u0131d\u0131r \u00e7\u00fcnk\u00fc evri\u015fim katmanlar\u0131, g\u00f6r\u00fcnt\u00fclerin yerel yap\u0131lar\u0131n\u0131 ve uzamsal ili\u015fkilerini koruyarak daha anlaml\u0131 ve verimli \u00f6zellikler \u00f6\u011frenir. Bu, daha y\u00fcksek s\u0131k\u0131\u015ft\u0131rma oranlar\u0131 ve daha iyi yeniden olu\u015fturma kalitesi anlam\u0131na gelir.<\/p>\n Seyrek otoenkoderler, gizli katmandaki n\u00f6ronlar\u0131n k\u00fc\u00e7\u00fck bir k\u0131sm\u0131n\u0131n ayn\u0131 anda aktif olmas\u0131n\u0131 zorlayarak \u00e7al\u0131\u015f\u0131r. Bu seyreklik k\u0131s\u0131tlamas\u0131, modelin daha kompakt ve verimli temsiller \u00f6\u011frenmesini sa\u011flar. Seyrek otoenkoderler, giri\u015f verisindeki temel \u00f6zelliklerin daha iyi yakalanmas\u0131na yard\u0131mc\u0131 olabilir ve a\u015f\u0131r\u0131 uydurmay\u0131 (overfitting) azaltabilir. G\u00f6r\u00fcnt\u00fc s\u0131k\u0131\u015ft\u0131rma ba\u011flam\u0131nda, bu, daha az bilgiyle daha zengin bir temsil elde edilmesine olanak tan\u0131yabilir.<\/p>\n G\u00fcr\u00fclt\u00fc giderici otoenkoderler, giri\u015f g\u00f6r\u00fcnt\u00fcs\u00fcne kas\u0131tl\u0131 olarak g\u00fcr\u00fclt\u00fc eklenerek e\u011fitilir ve bu g\u00fcr\u00fclt\u00fcl\u00fc giri\u015ften temiz, orijinal g\u00f6r\u00fcnt\u00fcy\u00fc yeniden olu\u015fturmaya \u00e7al\u0131\u015f\u0131r. Bu e\u011fitim stratejisi, modelin giri\u015f verisindeki temel \u00f6zellikleri g\u00fcr\u00fclt\u00fcden ay\u0131rmay\u0131 \u00f6\u011frenmesini sa\u011flar. G\u00f6r\u00fcnt\u00fc s\u0131k\u0131\u015ft\u0131rmada, bu t\u00fcr otoenkoderler, daha sa\u011flam (robust) ve g\u00fcr\u00fclt\u00fcye daha az duyarl\u0131 \u00f6zellik temsilleri \u00f6\u011frenerek dolayl\u0131 yoldan fayda sa\u011flayabilir.<\/p>\n VAE’ler, otoenkoder mimarisinin \u00fcretken bir uzant\u0131s\u0131d\u0131r. Geleneksel otoenkoderlerin aksine, VAE’ler gizli uzay\u0131 belirli bir olas\u0131l\u0131ksal da\u011f\u0131l\u0131m (genellikle Gauss da\u011f\u0131l\u0131m\u0131) olarak temsil eder. Kodlay\u0131c\u0131, giri\u015f verisi i\u00e7in bu da\u011f\u0131l\u0131m\u0131n ortalama ve varyans parametrelerini \u00f6\u011frenir. Kod \u00e7\u00f6z\u00fcc\u00fc ise bu da\u011f\u0131l\u0131mdan \u00f6rneklenen bir noktay\u0131 kullanarak g\u00f6r\u00fcnt\u00fcy\u00fc yeniden olu\u015fturur. VAE’ler, gizli uzay\u0131n daha p\u00fcr\u00fczs\u00fcz ve anlaml\u0131 olmas\u0131n\u0131 sa\u011flar, bu da yeni ve ger\u00e7ek\u00e7i g\u00f6r\u00fcnt\u00fcler \u00fcretme yetene\u011fi kazand\u0131r\u0131r. G\u00f6r\u00fcnt\u00fc s\u0131k\u0131\u015ft\u0131rma i\u00e7in, VAE’ler, gizli uzayda daha anlaml\u0131 ve yap\u0131sal olarak zengin temsiller \u00f6\u011frenerek potansiyel olarak daha iyi s\u0131k\u0131\u015ft\u0131rma performans\u0131 sunabilirler, ancak genellikle daha karma\u015f\u0131k bir e\u011fitim s\u00fcreci gerektirirler.<\/p>\n \u015eimdi, Keras kullanarak evri\u015fimsel bir otoenkoderin g\u00f6r\u00fcnt\u00fc s\u0131k\u0131\u015ft\u0131rma amac\u0131yla nas\u0131l in\u015fa edilece\u011fini, e\u011fitilece\u011fini ve de\u011ferlendirilece\u011fini ad\u0131m ad\u0131m a\u00e7\u0131klayal\u0131m.<\/p>\n G\u00f6r\u00fcnt\u00fc s\u0131k\u0131\u015ft\u0131rma uygulamalar\u0131 i\u00e7in genellikle b\u00fcy\u00fck ve \u00e7e\u015fitli g\u00f6r\u00fcnt\u00fc veri k\u00fcmeleri tercih edilir. Ancak, bir ba\u015flang\u0131\u00e7 ve kavramsal ispat i\u00e7in MNIST (el yaz\u0131s\u0131 rakamlar) veya CIFAR-10 (k\u00fc\u00e7\u00fck renkli g\u00f6r\u00fcnt\u00fcler) gibi daha k\u00fc\u00e7\u00fck veri k\u00fcmeleri de kullan\u0131labilir. Bu \u00f6rnekte, basitli\u011fi nedeniyle MNIST veri k\u00fcmesini kullanaca\u011f\u0131z. MNIST, 28×28 piksel boyutunda gri tonlamal\u0131 el yaz\u0131s\u0131 rakam g\u00f6r\u00fcnt\u00fclerinden olu\u015fur.<\/p>\n Veri \u00f6n i\u015fleme, derin \u00f6\u011frenme modellerinin e\u011fitimi i\u00e7in kritik bir ad\u0131md\u0131r. Evri\u015fimsel otoenkoder modelini Keras’\u0131n Functional API’sini kullanarak olu\u015fturaca\u011f\u0131z. Bu API, daha esnek ve karma\u015f\u0131k model mimarileri tan\u0131mlamam\u0131za olanak tan\u0131r.<\/p>\n Yukar\u0131daki kodda, Modeli e\u011fitmeden \u00f6nce, bir optimize edici (optimizer) ve bir kay\u0131p fonksiyonu (loss function) belirtmemiz gerekir. Modeli, e\u011fitim verileri \u00fczerinde belirli bir say\u0131da epok (epoch) ve y\u0131\u011f\u0131n boyutu (batch size) kullanarak e\u011fitebiliriz.<\/p>\n Burada E\u011fitim tamamland\u0131ktan sonra, otoenkoder modelini kullanarak g\u00f6r\u00fcnt\u00fcleri s\u0131k\u0131\u015ft\u0131rabilir ve yeniden olu\u015fturabiliriz.<\/p>\n Modelin performans\u0131n\u0131 de\u011ferlendirmek i\u00e7in orijinal ve yeniden olu\u015fturulmu\u015f g\u00f6r\u00fcnt\u00fcler aras\u0131nda g\u00f6rsel kar\u015f\u0131la\u015ft\u0131rmalar yapabiliriz. Ayr\u0131ca, PSNR ve SSIM gibi nicel metrikleri de hesaplayabiliriz.<\/p>\n Bu \u00e7\u0131kt\u0131lar, otoenkoderin s\u0131k\u0131\u015ft\u0131rma sonras\u0131 g\u00f6r\u00fcnt\u00fcy\u00fc ne kadar iyi yeniden olu\u015fturabildi\u011fini g\u00f6sterir. S\u0131k\u0131\u015ft\u0131rma oran\u0131 ise gizli uzay\u0131n boyutuna ve bu uzay\u0131n temsil etti\u011fi bilgi miktar\u0131na ba\u011fl\u0131d\u0131r. \u00d6rne\u011fin, (7, 7, 32) boyutundaki bir gizli uzay, orijinal (28, 28, 1) g\u00f6r\u00fcnt\u00fcs\u00fcne k\u0131yasla \u00f6nemli \u00f6l\u00e7\u00fcde daha az veri noktas\u0131 i\u00e7erir.<\/p>\n Otoenkoderlerin g\u00f6r\u00fcnt\u00fc s\u0131k\u0131\u015ft\u0131rma alan\u0131na getirdi\u011fi yenilikler ve potansiyeller, geleneksel y\u00f6ntemlere k\u0131yasla baz\u0131 belirgin avantajlar sunarken, beraberinde baz\u0131 zorluklar\u0131 da getirir.<\/p>\n * Veriye \u00d6zg\u00fc S\u0131k\u0131\u015ft\u0131rma:<\/strong> Otoenkoderler, e\u011fitildikleri veri k\u00fcmesindeki g\u00f6r\u00fcnt\u00fclerin \u00f6zelliklerini \u00f6\u011frenir. Bu, belirli bir t\u00fcr g\u00f6r\u00fcnt\u00fc (\u00f6rne\u011fin, t\u0131bbi g\u00f6r\u00fcnt\u00fcler, uydu g\u00f6r\u00fcnt\u00fcleri) i\u00e7in geleneksel y\u00f6ntemlerden daha optimize ve verimli s\u0131k\u0131\u015ft\u0131rma algoritmalar\u0131 olu\u015fturma potansiyeli sunar. Geleneksel algoritmalar genellikle genel ama\u00e7l\u0131d\u0131r. * E\u011fitim S\u00fcresi ve Hesaplama Maliyeti:<\/strong> Derin \u00f6\u011frenme modellerinin e\u011fitimi, \u00f6zellikle b\u00fcy\u00fck veri k\u00fcmeleri ve karma\u015f\u0131k mimariler i\u00e7in \u00f6nemli miktarda i\u015flem g\u00fcc\u00fc (GPU\/TPU) ve zaman gerektirir. Bu, modelin geli\u015ftirme ve da\u011f\u0131t\u0131m maliyetlerini art\u0131rabilir. Otoenkoder tabanl\u0131 g\u00f6r\u00fcnt\u00fc s\u0131k\u0131\u015ft\u0131rma alan\u0131, derin \u00f6\u011frenmenin s\u00fcrekli geli\u015fimiyle birlikte h\u0131zla ilerlemektedir. Gelecekte bu alanda beklenen baz\u0131 geli\u015fmeler ve ara\u015ft\u0131rma y\u00f6nleri \u015funlard\u0131r:<\/p>\n * \u00dcretken \u00c7eki\u015fmeli A\u011flar (GANs) ile S\u0131k\u0131\u015ft\u0131rma:<\/strong> GAN’lar, ger\u00e7ek\u00e7i g\u00f6r\u00fcnt\u00fcler \u00fcretme yetenekleriyle bilinir. S\u0131k\u0131\u015ft\u0131rma ba\u011flam\u0131nda, bir \u00fcretici a\u011f\u0131 s\u0131k\u0131\u015ft\u0131r\u0131lm\u0131\u015f temsilden g\u00f6r\u00fcnt\u00fcler olu\u015ftururken, bir ay\u0131rt edici a\u011f bu g\u00f6r\u00fcnt\u00fcleri ger\u00e7ek g\u00f6r\u00fcnt\u00fclerden ay\u0131rmaya \u00e7al\u0131\u015fabilir. Bu, daha y\u00fcksek g\u00f6rsel kaliteye sahip yeniden olu\u015fturulmu\u015f g\u00f6r\u00fcnt\u00fcler elde etme potansiyeli sunar. Otoenkoderler, derin \u00f6\u011frenmenin g\u00f6r\u00fcnt\u00fc s\u0131k\u0131\u015ft\u0131rma alan\u0131na getirdi\u011fi en umut verici yakla\u015f\u0131mlardan biridir. Kendi kendine \u00f6\u011frenen bu sinir a\u011flar\u0131, g\u00f6r\u00fcnt\u00fclerin temel \u00f6zelliklerini ve desenlerini ke\u015ffederek, onlar\u0131 d\u00fc\u015f\u00fck boyutlu ve verimli bir gizli uzayda temsil etme yetene\u011fine sahiptir. \u00d6zellikle evri\u015fimsel otoenkoderler, g\u00f6r\u00fcnt\u00fclerin uzamsal yap\u0131s\u0131n\u0131 koruyarak ve hiyerar\u015fik \u00f6zellikleri \u00f6\u011frenerek, geleneksel s\u0131k\u0131\u015ft\u0131rma y\u00f6ntemlerinin \u00f6tesinde performans sergileyebilirler.<\/p>\n Keras gibi kullan\u0131c\u0131 dostu k\u00fct\u00fcphaneler sayesinde, otoenkoder tabanl\u0131 g\u00f6r\u00fcnt\u00fc s\u0131k\u0131\u015ft\u0131rma modellerini olu\u015fturmak, e\u011fitmek ve de\u011ferlendirmek giderek daha eri\u015filebilir hale gelmi\u015ftir. Bu makalede ele ald\u0131\u011f\u0131m\u0131z ad\u0131mlar, basit bir evri\u015fimsel otoenkoderin MNIST veri k\u00fcmesi \u00fczerinde nas\u0131l \u00e7al\u0131\u015ft\u0131\u011f\u0131n\u0131 g\u00f6stermi\u015ftir. Elde edilen sonu\u00e7lar, otoenkoderlerin orijinal g\u00f6r\u00fcnt\u00fcy\u00fc \u00f6nemli \u00f6l\u00e7\u00fcde s\u0131k\u0131\u015ft\u0131rarak kabul edilebilir bir kalitede yeniden olu\u015fturabildi\u011fini ortaya koymaktad\u0131r.<\/p>\n Elbette, otoenkoder tabanl\u0131 s\u0131k\u0131\u015ft\u0131rman\u0131n hen\u00fcz geleneksel standartlar kadar yayg\u0131n veya optimize olmad\u0131\u011f\u0131 baz\u0131 zorluklar\u0131 bulunmaktad\u0131r. Y\u00fcksek hesaplama maliyetleri, genelleme sorunlar\u0131 ve standartla\u015fma eksikli\u011fi gibi konular, bu teknolojinin daha geni\u015f \u00e7apta benimsenmesi i\u00e7in a\u015f\u0131lmas\u0131 gereken engellerdir. Ancak, devam eden ara\u015ft\u0131rmalar ve derin \u00f6\u011frenme alan\u0131ndaki h\u0131zl\u0131 ilerlemeler, bu zorluklar\u0131n \u00fcstesinden gelmek ve otoenkoderleri gelece\u011fin g\u00f6r\u00fcnt\u00fc ve video s\u0131k\u0131\u015ft\u0131rma standartlar\u0131n\u0131n ayr\u0131lmaz bir par\u00e7as\u0131 haline getirmek i\u00e7in b\u00fcy\u00fck bir potansiyel sunmaktad\u0131r.<\/p>\n \u00d6zetle, otoenkoderler, g\u00f6r\u00fcnt\u00fc s\u0131k\u0131\u015ft\u0131rman\u0131n gelece\u011finde \u00f6nemli bir rol oynayacak ve daha verimli, adaptif ve y\u00fcksek kaliteli s\u0131k\u0131\u015ft\u0131rma \u00e7\u00f6z\u00fcmlerinin geli\u015ftirilmesine katk\u0131da bulunacakt\u0131r. Derin \u00f6\u011frenmenin bu alandaki etkisi, dijital \u00e7a\u011f\u0131n artan veri hacmiyle ba\u015fa \u00e7\u0131kmak i\u00e7in kritik \u00f6neme sahiptir.<\/body><\/p>\n","protected":false},"excerpt":{"rendered":"Keras ile Otoenkoder Kullanarak G\u00f6r\u00fcnt\u00fc S\u0131k\u0131\u015ft\u0131rma\nG\u00f6r\u00fcnt\u00fc verileri, modern dijital \u00e7a\u011f\u0131n en yayg\u0131n ve hacimli i\u00e7erik t\u00fcrlerinden bir","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"csco_page_header_type":"","csco_page_load_nextpost":"","csco_page_subscribe_form":"","csco_page_contact_form":"","footnotes":""},"categories":[1],"tags":[],"class_list":{"0":"post-36201","1":"post","2":"type-post","3":"status-publish","4":"format-standard","6":"category-genel","7":"cs-entry","8":"cs-video-wrap"},"yoast_head":"\n
\n* Ortalama Kare Hata (Mean Squared Error – MSE):<\/strong> Orijinal g\u00f6r\u00fcnt\u00fc ile yeniden olu\u015fturulmu\u015f g\u00f6r\u00fcnt\u00fc aras\u0131ndaki piksel de\u011ferlerinin kare farklar\u0131n\u0131n ortalamas\u0131d\u0131r. D\u00fc\u015f\u00fck MSE, daha iyi kalite anlam\u0131na gelir.
\n* Tepe Sinyal G\u00fcr\u00fclt\u00fc Oran\u0131 (Peak Signal-to-Noise Ratio – PSNR):<\/strong> MSE’den t\u00fcretilen ve desibel (dB) cinsinden ifade edilen bir metriktir. Daha y\u00fcksek PSNR de\u011ferleri, daha iyi g\u00f6r\u00fcnt\u00fc kalitesini g\u00f6sterir. Genellikle 30 dB’nin \u00fczerindeki de\u011ferler kabul edilebilir olarak kabul edilir.
\n* Yap\u0131sal Benzerlik \u0130ndeksi (Structural Similarity Index Measure – SSIM):<\/strong> \u0130nsan g\u00f6rsel sisteminin alg\u0131sal \u00f6zelliklerini dikkate alarak g\u00f6r\u00fcnt\u00fc kalitesini de\u011ferlendiren daha geli\u015fmi\u015f bir metriktir. Parlakl\u0131k, kontrast ve yap\u0131sal bilgideki de\u011fi\u015fiklikleri \u00f6l\u00e7er. 0 ile 1 aras\u0131nda bir de\u011fer al\u0131r; 1’e yak\u0131n de\u011ferler daha y\u00fcksek benzerli\u011fi ve kaliteyi g\u00f6sterir.<\/p>\nOtoenkoderlere Giri\u015f<\/h2>\n
Temel Mimari: Kodlay\u0131c\u0131 ve Kod \u00c7\u00f6z\u00fcc\u00fc<\/h3>\n
\n1. Kodlay\u0131c\u0131 (Encoder):<\/strong> Giri\u015f verisini al\u0131r ve onu daha d\u00fc\u015f\u00fck boyutlu, s\u0131k\u0131\u015ft\u0131r\u0131lm\u0131\u015f bir temsile d\u00f6n\u00fc\u015ft\u00fcr\u00fcr. Bu s\u0131k\u0131\u015ft\u0131r\u0131lm\u0131\u015f temsile “gizli uzay” (latent space) veya “kod” (code) denir. Kodlay\u0131c\u0131, giri\u015f verisindeki en \u00f6nemli \u00f6zellikleri ve desenleri \u00f6\u011frenerek bu temsili olu\u015fturur. Genellikle, katman say\u0131s\u0131 ve n\u00f6ron say\u0131s\u0131 azalan bir yap\u0131ya sahiptir.
\n2. Kod \u00c7\u00f6z\u00fcc\u00fc (Decoder):<\/strong> Gizli uzaydaki s\u0131k\u0131\u015ft\u0131r\u0131lm\u0131\u015f temsili al\u0131r ve orijinal giri\u015f verisine benzer bir \u00e7\u0131k\u0131\u015f \u00fcretmeye \u00e7al\u0131\u015f\u0131r. Kod \u00e7\u00f6z\u00fcc\u00fc, kodlay\u0131c\u0131n\u0131n tersi bir yap\u0131ya sahiptir; yani katman say\u0131s\u0131 ve n\u00f6ron say\u0131s\u0131 artan bir yap\u0131ya sahiptir.<\/p>\nGizli Uzay (Latent Space) Kavram\u0131<\/h3>\n
Otoenkoder \u00c7e\u015fitleri ve G\u00f6r\u00fcnt\u00fc S\u0131k\u0131\u015ft\u0131rmadaki Rolleri<\/h2>\n
Basit Otoenkoderler<\/h3>\n
Evri\u015fimsel Otoenkoderler (Convolutional Autoencoders – CAE)<\/h3>\n
Conv2D<\/code> katman\u0131, MaxPooling2D<\/code> katman\u0131 ve aktivasyon fonksiyonlar\u0131ndan olu\u015fur. Conv2D<\/code> katmanlar\u0131, g\u00f6r\u00fcnt\u00fcdeki yerel desenleri (kenarlar, k\u00f6\u015feler, dokular) \u00f6\u011frenir. MaxPooling2D<\/code> katmanlar\u0131 ise uzamsal boyutlar\u0131 k\u00fc\u00e7\u00fclterek ve en \u00f6nemli \u00f6zellikleri koruyarak verinin boyutunu azalt\u0131r. Bu katmanlar, g\u00f6r\u00fcnt\u00fcy\u00fc giderek daha d\u00fc\u015f\u00fck boyutlu ve daha soyut bir temsile d\u00f6n\u00fc\u015ft\u00fcr\u00fcr.
\n* Gizli Uzay:<\/strong> Kodlay\u0131c\u0131n\u0131n \u00e7\u0131k\u0131\u015f\u0131, genellikle son MaxPooling2D<\/code> katman\u0131n\u0131n veya bir Dense<\/code> katman\u0131n\u0131n \u00e7\u0131kt\u0131s\u0131 olan, s\u0131k\u0131\u015ft\u0131r\u0131lm\u0131\u015f gizli uzay\u0131 olu\u015fturur. Bu, orijinal g\u00f6r\u00fcnt\u00fcn\u00fcn d\u00fc\u015f\u00fck boyutlu, anlaml\u0131 bir temsilidir.
\n* Kod \u00c7\u00f6z\u00fcc\u00fc:<\/strong> Evri\u015fimsel kod \u00e7\u00f6z\u00fcc\u00fc, gizli uzaydaki temsili orijinal g\u00f6r\u00fcnt\u00fc boyutlar\u0131na geri d\u00f6nd\u00fcrmek i\u00e7in bir dizi Conv2DTranspose<\/code> (deconvolution veya kesirli ad\u0131ml\u0131 evri\u015fim) veya UpSampling2D<\/code> katmanlar\u0131 ile Conv2D<\/code> katmanlar\u0131n\u0131 kullan\u0131r. Conv2DTranspose<\/code> katmanlar\u0131, Conv2D<\/code> katmanlar\u0131n\u0131n tersi i\u015flevi g\u00f6rerek uzamsal boyutlar\u0131 art\u0131r\u0131r ve \u00f6zellik haritalar\u0131n\u0131 yeniden birle\u015ftirir. Son katman genellikle sigmoid<\/code> aktivasyon fonksiyonu ile 0-1 aral\u0131\u011f\u0131nda piksel de\u011ferleri \u00fcreten bir Conv2D<\/code> katman\u0131d\u0131r.<\/p>\nSeyrek Otoenkoderler (Sparse Autoencoders)<\/h3>\n
G\u00fcr\u00fclt\u00fc Giderici Otoenkoderler (Denoising Autoencoders)<\/h3>\n
Varyasyonel Otoenkoderler (Variational Autoencoders – VAE)<\/h3>\n
Keras ile Otoenkoder G\u00f6r\u00fcnt\u00fc S\u0131k\u0131\u015ft\u0131rma Uygulamas\u0131<\/h2>\n
Veri K\u00fcmesi Se\u00e7imi<\/h3>\n
Veri \u00d6n \u0130\u015fleme<\/h3>\n
\n1. G\u00f6r\u00fcnt\u00fclerin Y\u00fcklenmesi:<\/strong> Keras, mnist.load_data()<\/code> fonksiyonu ile veri k\u00fcmesini kolayca y\u00fcklememizi sa\u011flar.
\n2. Normalizasyon:<\/strong> Piksel de\u011ferleri genellikle 0-255 aral\u0131\u011f\u0131ndad\u0131r. Bu de\u011ferleri 0-1 aral\u0131\u011f\u0131na normalize etmek, modelin daha istikrarl\u0131 ve h\u0131zl\u0131 \u00f6\u011frenmesine yard\u0131mc\u0131 olur. Bu, her piksel de\u011ferini 255.0’e b\u00f6lerek yap\u0131l\u0131r.
\n3. Boyutland\u0131rma (Reshape):<\/strong> Evri\u015fimli katmanlar, giri\u015f olarak (y\u00fckseklik, geni\u015flik, kanal) format\u0131nda g\u00f6r\u00fcnt\u00fcler bekler. MNIST g\u00f6r\u00fcnt\u00fcleri gri tonlamal\u0131 oldu\u011fu i\u00e7in tek bir kanala sahiptir. Bu nedenle, 28×28 boyutundaki g\u00f6r\u00fcnt\u00fcleri (28, 28, 1) \u015fekline getirmemiz gerekir.<\/p>\nimport numpy as np\nfrom tensorflow import keras\nfrom tensorflow.keras import layers\n\nVeri k\u00fcmesini y\u00fckle<\/h2>\n(x_train, _), (x_test, _) = keras.datasets.mnist.load_data()\n\n
Piksel de\u011ferlerini 0-1 aral\u0131\u011f\u0131na normalize et<\/h2>\nx_train = x_train.astype('float32') \/ 255.\nx_test = x_test.astype('float32') \/ 255.\n\n
G\u00f6r\u00fcnt\u00fcleri (28, 28, 1) \u015fekline getir (gri tonlamal\u0131, 1 kanal)<\/h2>\nx_train = np.reshape(x_train, (len(x_train), 28, 28, 1))\nx_test = np.reshape(x_test, (len(x_test), 28, 28, 1))<\/code><\/pre>\n
Model Mimarisi Olu\u015fturma<\/h3>\n
# Giri\u015f g\u00f6r\u00fcnt\u00fcs\u00fc boyutu\ninput_img = keras.Input(shape=(28, 28, 1))\n\n--- Kodlay\u0131c\u0131 (Encoder) ---<\/h2>\n
1. Evri\u015fim katman\u0131: 32 filtre, 3x3 kernel, relu aktivasyon, padding='same' ile boyut korunur<\/h2>\nx = layers.Conv2D(32, (3, 3), activation='relu', padding='same')(input_img)\n
MaxPooling: Uzamsal boyutlar\u0131 yar\u0131ya indirir (28x28 -> 14x14)<\/h2>\nx = layers.MaxPooling2D((2, 2), padding='same')(x)\n
2. Evri\u015fim katman\u0131: 32 filtre, 3x3 kernel, relu aktivasyon<\/h2>\nx = layers.Conv2D(32, (3, 3), activation='relu', padding='same')(x)\n
MaxPooling: Uzamsal boyutlar\u0131 yar\u0131ya indirir (14x14 -> 7x7)<\/h2>\nencoded = layers.MaxPooling2D((2, 2), padding='same')(x)\n\n
encoded (gizli uzay) \u015fimdi (7, 7, 32) boyutunda bir tens\u00f6rd\u00fcr.<\/h2>\n
Bu, orijinal 28x28x1 g\u00f6r\u00fcnt\u00fcn\u00fcn s\u0131k\u0131\u015ft\u0131r\u0131lm\u0131\u015f temsilidir.<\/h2>\n
S\u0131k\u0131\u015ft\u0131rma oran\u0131: (2828<\/em>1) \/ (77<\/em>32) = 784 \/ 1568 = 0.5 (Bu tam olarak s\u0131k\u0131\u015ft\u0131rma oran\u0131 de\u011fil,<\/h2>\n
ancak boyut k\u00fc\u00e7\u00fclmesini g\u00f6sterir. Bit baz\u0131nda s\u0131k\u0131\u015ft\u0131rma i\u00e7in daha derin analiz gerekir.)<\/h2>\n\n
--- Kod \u00c7\u00f6z\u00fcc\u00fc (Decoder) ---<\/h2>\n
Upsampling: Uzamsal boyutlar\u0131 iki kat\u0131na \u00e7\u0131kar\u0131r (7x7 -> 14x14)<\/h2>\nx = layers.Conv2D(32, (3, 3), activation='relu', padding='same')(encoded)\nx = layers.UpSampling2D((2, 2))(x)\n
1. Evri\u015fim katman\u0131: 32 filtre, 3x3 kernel, relu aktivasyon<\/h2>\nx = layers.Conv2D(32, (3, 3), activation='relu', padding='same')(x)\n
Upsampling: Uzamsal boyutlar\u0131 iki kat\u0131na \u00e7\u0131kar\u0131r (14x14 -> 28x28)<\/h2>\nx = layers.UpSampling2D((2, 2))(x)\n
\u00c7\u0131k\u0131\u015f katman\u0131: 1 filtre (gri tonlamal\u0131 g\u00f6r\u00fcnt\u00fc i\u00e7in), 3x3 kernel, sigmoid aktivasyon<\/h2>\n
Sigmoid, piksel de\u011ferlerini 0-1 aral\u0131\u011f\u0131na s\u0131k\u0131\u015ft\u0131r\u0131r.<\/h2>\ndecoded = layers.Conv2D(1, (3, 3), activation='sigmoid', padding='same')(x)\n\n
Otoenkoder modelini olu\u015ftur<\/h2>\nautoencoder = keras.Model(input_img, decoded)\n\n
Sadece kodlay\u0131c\u0131 k\u0131sm\u0131n\u0131 i\u00e7eren modeli olu\u015ftur (s\u0131k\u0131\u015ft\u0131rma i\u00e7in)<\/h2>\nencoder = keras.Model(input_img, encoded)\n\n
Sadece kod \u00e7\u00f6z\u00fcc\u00fc k\u0131sm\u0131n\u0131 i\u00e7eren modeli olu\u015ftur (yeniden olu\u015fturma i\u00e7in)<\/h2>\n
Kod \u00e7\u00f6z\u00fcc\u00fc girdisi, kodlay\u0131c\u0131n\u0131n \u00e7\u0131kt\u0131s\u0131n\u0131n \u015fekli olmal\u0131d\u0131r: (7, 7, 32)<\/h2>\nencoded_input = keras.Input(shape=(7, 7, 32))\ndecoder_layer = autoencoder.layers[-7](encoded_input) # \u0130lk UpSampling katman\u0131\ndecoder_layer = autoencoder.layers[-6](decoder_layer) # \u0130lk Conv2D katman\u0131\ndecoder_layer = autoencoder.layers[-5](decoder_layer) # \u0130kinci UpSampling katman\u0131\ndecoder_layer = autoencoder.layers[-4](decoder_layer) # \u0130kinci Conv2D katman\u0131\ndecoder_layer = autoencoder.layers[-3](decoder_layer) # Son Conv2D katman\u0131 (output)\ndecoder = keras.Model(encoded_input, decoder_layer)<\/code><\/pre>\n
autoencoder.layers[-X]<\/code> kullanarak kod \u00e7\u00f6z\u00fcc\u00fcy\u00fc olu\u015fturmak biraz karma\u015f\u0131k olabilir. Daha temiz bir yol, kod \u00e7\u00f6z\u00fcc\u00fc katmanlar\u0131n\u0131 ayr\u0131 ayr\u0131 tan\u0131mlay\u0131p sonra birle\u015ftirmektir, ancak bu \u00f6rnekte mevcut modelin katmanlar\u0131n\u0131 yeniden kullanma mant\u0131\u011f\u0131 g\u00f6sterilmi\u015ftir.<\/p>\nModelin Derlenmesi<\/h3>\n
\n* Optimize Edici:<\/strong> Adam<\/code> optimize edici, \u00e7o\u011fu derin \u00f6\u011frenme g\u00f6revi i\u00e7in iyi bir ba\u015flang\u0131\u00e7 noktas\u0131d\u0131r.
\n* Kay\u0131p Fonksiyonu:<\/strong> G\u00f6r\u00fcnt\u00fc pikselleri 0-1 aral\u0131\u011f\u0131na normalize edildi\u011fi ve \u00e7\u0131k\u0131\u015f katman\u0131 sigmoid<\/code> aktivasyonu kulland\u0131\u011f\u0131 i\u00e7in mean_squared_error<\/code> (MSE) veya binary_crossentropy<\/code> uygundur. MSE, piksel de\u011ferleri aras\u0131ndaki farklar\u0131 kareleyerek hata \u00f6l\u00e7er.<\/p>\nautoencoder.compile(optimizer='adam', loss='mean_squared_error')<\/code><\/pre>\nModelin E\u011fitimi<\/h3>\n
autoencoder.fit(x_train, x_train,\n epochs=50,\n batch_size=256,\n shuffle=True,\n validation_data=(x_test, x_test))<\/code><\/pre>\nx_train<\/code> hem giri\u015f hem de beklenen \u00e7\u0131k\u0131\u015f olarak kullan\u0131l\u0131r, \u00e7\u00fcnk\u00fc otoenkoderler denetimsiz \u00f6\u011frenme paradigmas\u0131nda \u00e7al\u0131\u015f\u0131r ve kendi girdilerini yeniden olu\u015fturmay\u0131 \u00f6\u011frenir. validation_data<\/code> ile modelin e\u011fitim s\u0131ras\u0131nda test verisi \u00fczerindeki performans\u0131n\u0131 izleyebiliriz.<\/p>\nS\u0131k\u0131\u015ft\u0131rma ve Yeniden Olu\u015fturma<\/h3>\n
# Test g\u00f6r\u00fcnt\u00fclerinden baz\u0131lar\u0131n\u0131 s\u0131k\u0131\u015ft\u0131r\nencoded_imgs = encoder.predict(x_test)\n\nS\u0131k\u0131\u015ft\u0131r\u0131lm\u0131\u015f temsillerden g\u00f6r\u00fcnt\u00fcleri yeniden olu\u015ftur<\/h2>\ndecoded_imgs = autoencoder.predict(x_test) # veya decoder.predict(encoded_imgs)<\/code><\/pre>\n
De\u011ferlendirme<\/h3>\n
import matplotlib.pyplot as plt\n\nn = 10 # G\u00f6sterilecek g\u00f6r\u00fcnt\u00fc say\u0131s\u0131\nplt.figure(figsize=(20, 4))\nfor i in range(n):\n # Orijinal g\u00f6r\u00fcnt\u00fcler\n ax = plt.subplot(2, n, i + 1)\n plt.imshow(x_test[i].reshape(28, 28))\n plt.gray()\n ax.get_xaxis().set_visible(False)\n ax.get_yaxis().set_visible(False)\n\n # Yeniden olu\u015fturulmu\u015f g\u00f6r\u00fcnt\u00fcler\n ax = plt.subplot(2, n, i + 1 + n)\n plt.imshow(decoded_imgs[i].reshape(28, 28))\n plt.gray()\n ax.get_xaxis().set_visible(False)\n ax.get_yaxis().set_visible(False)\nplt.show()\n\nPSNR ve SSIM hesaplamas\u0131 i\u00e7in scikit-image k\u00fct\u00fcphanesi kullan\u0131labilir<\/h2>\nfrom skimage.metrics import peak_signal_noise_ratio, structural_similarity\n\npsnr_values = []\nssim_values = []\n\nfor i in range(len(x_test)):\n original = x_test[i].reshape(28, 28)\n reconstructed = decoded_imgs[i].reshape(28, 28)\n \n psnr_values.append(peak_signal_noise_ratio(original, reconstructed, data_range=1.0))\n ssim_values.append(structural_similarity(original, reconstructed, data_range=1.0))\n\nprint(f\"Ortalama PSNR: {np.mean(psnr_values):.2f} dB\")\nprint(f\"Ortalama SSIM: {np.mean(ssim_values):.2f}\")<\/code><\/pre>\n
Otoenkoder Tabanl\u0131 S\u0131k\u0131\u015ft\u0131rman\u0131n Avantajlar\u0131 ve Dezavantajlar\u0131<\/h2>\n
Avantajlar<\/h3>\n
\n* Daha \u0130yi S\u0131k\u0131\u015ft\u0131rma Oranlar\u0131 Potansiyeli:<\/strong> Derin \u00f6\u011frenme modelleri, g\u00f6r\u00fcnt\u00fcdeki anlamsal bilgiyi daha iyi anlayarak ve gereksiz fazlal\u0131klar\u0131 daha etkili bir \u015fekilde ortadan kald\u0131rarak, geleneksel y\u00f6ntemlere g\u00f6re daha y\u00fcksek s\u0131k\u0131\u015ft\u0131rma oranlar\u0131 elde etme potansiyeline sahiptir.
\n* G\u00fcr\u00fclt\u00fc Azaltma ve \u00d6zellik \u00d6\u011frenimi:<\/strong> Otoenkoderler, s\u0131k\u0131\u015ft\u0131rma g\u00f6revinin yan\u0131 s\u0131ra, g\u00fcr\u00fclt\u00fc azaltma ve g\u00f6r\u00fcnt\u00fclerden y\u00fcksek seviyeli \u00f6zellikler \u00f6\u011frenme gibi ek faydalar da sa\u011flayabilir. \u00d6zellikle g\u00fcr\u00fclt\u00fc giderici otoenkoderler, s\u0131k\u0131\u015ft\u0131rma s\u0131ras\u0131nda g\u00fcr\u00fclt\u00fcy\u00fc filtreleyerek daha temiz yeniden olu\u015fturulmu\u015f g\u00f6r\u00fcnt\u00fcler \u00fcretebilir.
\n* Esneklik:<\/strong> Keras gibi k\u00fct\u00fcphanelerle farkl\u0131 otoenkoder mimarileri, kay\u0131p fonksiyonlar\u0131 ve optimize ediciler denemek kolayd\u0131r. Bu esneklik, belirli uygulama gereksinimlerine g\u00f6re \u00f6zel s\u0131k\u0131\u015ft\u0131rma \u00e7\u00f6z\u00fcmleri geli\u015ftirmeye olanak tan\u0131r.
\n* Anlamsal S\u0131k\u0131\u015ft\u0131rma:<\/strong> Otoenkoderler, piksel d\u00fczeyindeki benzerlikten ziyade g\u00f6r\u00fcnt\u00fcn\u00fcn anlamsal i\u00e7eri\u011fini korumaya odaklanabilir. Bu, \u00e7ok y\u00fcksek s\u0131k\u0131\u015ft\u0131rma oranlar\u0131nda bile g\u00f6rsel olarak kabul edilebilir sonu\u00e7lar elde edilmesini sa\u011flayabilir.<\/p>\nDezavantajlar<\/h3>\n
\n* Model Boyutu:<\/strong> E\u011fitilmi\u015f otoenkoder modelleri, \u00f6zellikle \u00e7ok say\u0131da katman ve n\u00f6ron i\u00e7eriyorsa, kendileri de \u00f6nemli depolama alan\u0131 kaplayabilir. Bu, s\u0131k\u0131\u015ft\u0131rma modelinin kendisinin depolanmas\u0131 ve da\u011f\u0131t\u0131lmas\u0131 gerekti\u011finde bir sorun te\u015fkil edebilir.
\n* Genelleme Yetene\u011fi:<\/strong> Otoenkoderler, e\u011fitildikleri veri setine \u00f6zg\u00fcd\u00fcr. E\u011fitilmedikleri veya \u00e7ok farkl\u0131 \u00f6zelliklere sahip g\u00f6r\u00fcnt\u00fcler \u00fczerinde k\u00f6t\u00fc performans g\u00f6sterebilirler. Bu, geni\u015f bir g\u00f6r\u00fcnt\u00fc yelpazesini s\u0131k\u0131\u015ft\u0131rmak i\u00e7in genel ama\u00e7l\u0131 bir otoenkoder olu\u015fturmay\u0131 zorla\u015ft\u0131r\u0131r.
\n* Kay\u0131pl\u0131 S\u0131k\u0131\u015ft\u0131rma Do\u011fas\u0131:<\/strong> \u00c7o\u011fu otoenkoder tabanl\u0131 s\u0131k\u0131\u015ft\u0131rma, kay\u0131pl\u0131 bir s\u00fcre\u00e7tir. Yani, yeniden olu\u015fturulan g\u00f6r\u00fcnt\u00fc orijinal g\u00f6r\u00fcnt\u00fcden bir miktar farkl\u0131 olacakt\u0131r. Kritik uygulamalarda (t\u0131bbi g\u00f6r\u00fcnt\u00fcleme gibi) bu bilgi kayb\u0131 kabul edilemez olabilir.
\n* Standardizasyon Eksikli\u011fi:<\/strong> Geleneksel s\u0131k\u0131\u015ft\u0131rma formatlar\u0131 (JPEG, PNG) standartla\u015ft\u0131r\u0131lm\u0131\u015f ve geni\u015f \u00e7apta desteklenirken, otoenkoder tabanl\u0131 s\u0131k\u0131\u015ft\u0131rma i\u00e7in hen\u00fcz evrensel bir standart bulunmamaktad\u0131r. Bu, farkl\u0131 modeller aras\u0131nda uyumluluk sorunlar\u0131na yol a\u00e7abilir.
\n* S\u0131k\u0131\u015ft\u0131rma Oran\u0131n\u0131n Belirlenmesi:<\/strong> Geleneksel s\u0131k\u0131\u015ft\u0131rmada s\u0131k\u0131\u015ft\u0131rma oran\u0131 do\u011frudan kontrol edilebilirken, otoenkoderlerde bu, gizli uzay\u0131n boyutu, mimari ve e\u011fitim s\u00fcreci gibi bir\u00e7ok fakt\u00f6re ba\u011fl\u0131d\u0131r. Optimum s\u0131k\u0131\u015ft\u0131rma oran\u0131 ve kalite dengesini bulmak deneysel bir s\u00fcre\u00e7tir.<\/p>\nGelecek Perspektifleri ve Geli\u015fmeler<\/h2>\n
\n* Alg\u0131sal Kay\u0131p Fonksiyonlar\u0131 (Perceptual Loss):<\/strong> Geleneksel MSE gibi kay\u0131p fonksiyonlar\u0131 piksel d\u00fczeyindeki farkl\u0131l\u0131klara odaklan\u0131rken, alg\u0131sal kay\u0131p fonksiyonlar\u0131 insan g\u00f6rsel sisteminin alg\u0131sal \u00f6zelliklerini daha iyi taklit eder. Bu, daha y\u00fcksek s\u0131k\u0131\u015ft\u0131rma oranlar\u0131nda bile insan g\u00f6z\u00fc i\u00e7in daha ho\u015f ve ger\u00e7ek\u00e7i g\u00f6r\u00fcnen g\u00f6r\u00fcnt\u00fcler \u00fcretilmesini sa\u011flayabilir.
\n* Video S\u0131k\u0131\u015ft\u0131rmaya Uygulanmas\u0131:<\/strong> G\u00f6r\u00fcnt\u00fc s\u0131k\u0131\u015ft\u0131rma teknikleri, video s\u0131k\u0131\u015ft\u0131rma i\u00e7in de temel olu\u015fturur. Otoenkoderlerin zamansal ba\u011f\u0131ml\u0131l\u0131klar\u0131 \u00f6\u011frenen varyasyonlar\u0131 (\u00f6rne\u011fin, 3D evri\u015fimli otoenkoderler veya Recurrent Neural Network (RNN) tabanl\u0131 yap\u0131lar), video s\u0131k\u0131\u015ft\u0131rma performans\u0131n\u0131 \u00f6nemli \u00f6l\u00e7\u00fcde art\u0131rabilir.
\n* Daha Verimli Mimariler:<\/strong> Transformer tabanl\u0131 mimariler veya daha hafif evri\u015fimli a\u011flar gibi yeni sinir a\u011f\u0131 mimarileri, otoenkoderlerin hesaplama verimlili\u011fini ve s\u0131k\u0131\u015ft\u0131rma performans\u0131n\u0131 art\u0131rabilir. \u00d6zellikle mobil cihazlar ve g\u00f6m\u00fcl\u00fc sistemler i\u00e7in d\u00fc\u015f\u00fck kaynak t\u00fcketimli s\u0131k\u0131\u015ft\u0131rma \u00e7\u00f6z\u00fcmleri geli\u015ftirmek \u00f6nemlidir.
\n* Uyarlanabilir S\u0131k\u0131\u015ft\u0131rma:<\/strong> G\u00f6r\u00fcnt\u00fcn\u00fcn i\u00e7eri\u011fine ve karma\u015f\u0131kl\u0131\u011f\u0131na g\u00f6re s\u0131k\u0131\u015ft\u0131rma parametrelerini dinamik olarak ayarlayabilen otoenkoderler, farkl\u0131 b\u00f6lgelerde farkl\u0131 s\u0131k\u0131\u015ft\u0131rma oranlar\u0131 uygulayarak genel kaliteyi optimize edebilir.
\n* Kuantize Edilmi\u015f Gizli Uzaylar:<\/strong> Gizli uzay\u0131n kuantize edilmesi (yani, s\u00fcrekli de\u011ferlerin ayr\u0131k de\u011ferlere d\u00f6n\u00fc\u015ft\u00fcr\u00fclmesi), s\u0131k\u0131\u015ft\u0131r\u0131lm\u0131\u015f verinin daha da k\u00fc\u00e7\u00fck boyutlarda depolanmas\u0131na olanak tan\u0131r ve standart s\u0131k\u0131\u015ft\u0131rma algoritmalar\u0131yla daha kolay entegrasyon sa\u011flayabilir.<\/p>\nSonu\u00e7<\/h2>\n