Nesne tespiti (object detection), bilgisayar g\u00f6r\u00fc\u015f\u00fcn\u00fcn temel ve en kritik g\u00f6revlerinden biridir. Bir g\u00f6r\u00fcnt\u00fc veya video i\u00e7indeki nesnelerin konumunu (s\u0131n\u0131r kutular\u0131 ile) ve t\u00fcr\u00fcn\u00fc (s\u0131n\u0131f etiketleri ile) belirlemeyi ama\u00e7lar. Bu teknoloji, otonom ara\u00e7lardan g\u00fcvenlik kameralar\u0131na, t\u0131bbi g\u00f6r\u00fcnt\u00fclemeden end\u00fcstriyel otomasyona kadar say\u0131s\u0131z alanda devrim niteli\u011finde uygulamalara yol a\u00e7m\u0131\u015ft\u0131r. \u0130nsan-bilgisayar etkile\u015fimini, ak\u0131ll\u0131 \u015fehirleri ve daha bir\u00e7ok yenili\u011fi m\u00fcmk\u00fcn k\u0131lan nesne tespiti, yapay zeka ara\u015ft\u0131rmalar\u0131n\u0131n en dinamik alanlar\u0131ndan biridir.<\/p>\n
Nesne tespiti alan\u0131ndaki ilk yakla\u015f\u0131mlar, genellikle el yap\u0131m\u0131 \u00f6zellik \u00e7\u0131kar\u0131c\u0131lar (\u00f6rn. SIFT, HOG) ve geleneksel s\u0131n\u0131fland\u0131r\u0131c\u0131lar (\u00f6rn. SVM) kullan\u0131larak geli\u015ftirilmi\u015ftir. Bu y\u00f6ntemler, belirli ko\u015fullar alt\u0131nda ba\u015far\u0131l\u0131 olsalar da, karma\u015f\u0131k arka planlar, nesne varyasyonlar\u0131 ve ger\u00e7ek zamanl\u0131 performans gereksinimleri kar\u015f\u0131s\u0131nda yetersiz kalm\u0131\u015flard\u0131r. 2012 y\u0131l\u0131nda AlexNet’in ImageNet yar\u0131\u015fmas\u0131n\u0131 kazanmas\u0131yla derin \u00f6\u011frenmenin y\u00fckseli\u015fi, bilgisayar g\u00f6r\u00fc\u015f\u00fc alan\u0131nda bir d\u00f6n\u00fcm noktas\u0131 olmu\u015f ve nesne tespiti paradigmalar\u0131n\u0131 k\u00f6kten de\u011fi\u015ftirmi\u015ftir. Evri\u015fimsel Sinir A\u011flar\u0131 (Convolutional Neural Networks – CNN), g\u00f6r\u00fcnt\u00fclerden hiyerar\u015fik ve anlamsal \u00f6zellikler \u00f6\u011frenme yetenekleriyle, nesne tespiti modellerinin performans\u0131n\u0131 benzeri g\u00f6r\u00fclmemi\u015f seviyelere ta\u015f\u0131m\u0131\u015ft\u0131r.<\/p>\n
Derin \u00f6\u011frenme tabanl\u0131 nesne tespiti modelleri genellikle iki ana kategoriye ayr\u0131l\u0131r: Bu makalede, iki a\u015famal\u0131 tespit\u00e7ilerin en ba\u015far\u0131l\u0131 ve etkili temsilcilerinden biri olan Faster R-CNN mimarisine odaklanaca\u011f\u0131z. Faster R-CNN, kendisinden \u00f6nceki R-CNN ve Fast R-CNN modellerinin eksikliklerini gidererek, nesne tespiti alan\u0131nda \u00f6nemli bir ilerleme kaydetmi\u015f ve modern nesne tespit sistemlerinin temelini atm\u0131\u015ft\u0131r.<\/p>\n Faster R-CNN’i tam olarak anlamak i\u00e7in, evrimsel yolculu\u011funun ba\u015flang\u0131c\u0131 olan R-CNN ve Fast R-CNN modellerini incelemek faydal\u0131d\u0131r. Bu modeller, derin \u00f6\u011frenmenin nesne tespiti alan\u0131na nas\u0131l entegre edildi\u011fini ve zamanla nas\u0131l daha verimli hale getirildi\u011fini g\u00f6stermektedir.<\/p>\n R-CNN, Ross Girshick ve ekibi taraf\u0131ndan 2014 y\u0131l\u0131nda tan\u0131t\u0131lan, derin \u00f6\u011frenmeyi nesne tespiti ile birle\u015ftiren ilk \u00f6nemli modeldir. R-CNN’in temel \u00e7al\u0131\u015fma prensibi \u015fu ad\u0131mlardan olu\u015fur: R-CNN, derin \u00f6\u011frenmenin nesne tespiti do\u011frulu\u011funu \u00f6nemli \u00f6l\u00e7\u00fcde art\u0131rd\u0131\u011f\u0131n\u0131 g\u00f6sterse de, ciddi dezavantajlara sahipti: R-CNN’in eksikliklerini gidermek amac\u0131yla, Ross Girshick 2015 y\u0131l\u0131nda Fast R-CNN’i tan\u0131tt\u0131. Fast R-CNN, temel olarak R-CNN’in h\u0131z sorununu \u00e7\u00f6zmeyi ama\u00e7lad\u0131 ve bunu ROI Havuzlama (Region of Interest Pooling) katman\u0131n\u0131 kullanarak ba\u015fard\u0131. Fast R-CNN’in getirdi\u011fi yenilikler: Fast R-CNN, R-CNN’e g\u00f6re 9 kat daha h\u0131zl\u0131 e\u011fitim ve 213 kat daha h\u0131zl\u0131 test s\u00fcresi sunarak \u00f6nemli bir ilerleme kaydetti. Ancak, hala d\u0131\u015far\u0131dan bir b\u00f6lge \u00f6neri algoritmas\u0131na (Se\u00e7ici Arama) ba\u011f\u0131ml\u0131 olmas\u0131, modelin tamamen u\u00e7tan uca (end-to-end) bir derin \u00f6\u011frenme modeli olmas\u0131n\u0131 engelliyordu ve b\u00f6lge \u00f6neri s\u00fcreci hala bir darbo\u011faz olu\u015fturuyordu.<\/p>\n Fast R-CNN’in b\u00f6lge \u00f6neri a\u015famas\u0131ndaki ba\u011f\u0131ml\u0131l\u0131\u011f\u0131n\u0131 ortadan kald\u0131rmak amac\u0131yla, Shaoqing Ren ve ekibi 2015 y\u0131l\u0131nda Faster R-CNN’i tan\u0131tt\u0131. Faster R-CNN, b\u00f6lge \u00f6neri s\u00fcrecini de derin \u00f6\u011frenme a\u011f\u0131na entegre ederek, nesne tespiti i\u00e7in tamamen u\u00e7tan uca bir derin \u00f6\u011frenme mimarisi sunmu\u015ftur. Bu, modelin hem daha h\u0131zl\u0131 hem de daha do\u011fru olmas\u0131n\u0131 sa\u011flam\u0131\u015ft\u0131r.<\/p>\n Faster R-CNN mimarisi, temel olarak d\u00f6rt ana bile\u015fenden olu\u015fur: Bu d\u00f6rt bile\u015fen, tek bir a\u011fda birle\u015ftirilerek u\u00e7tan uca \u00f6\u011frenme ve \u00e7\u0131kar\u0131m sa\u011flar.<\/p>\n RPN, Faster R-CNN’in kalbidir ve modelin en b\u00fcy\u00fck yenili\u011fini temsil eder. RPN, payla\u015f\u0131lan evri\u015fimsel \u00f6zellik haritas\u0131 \u00fczerinde \u00e7al\u0131\u015farak nesne i\u00e7erme olas\u0131l\u0131\u011f\u0131 y\u00fcksek olan b\u00f6lgeleri do\u011frudan \u00f6\u011frenir.<\/p>\n RPN’in temelinde “anchor kutular\u0131” (anchor boxes) kavram\u0131 yer al\u0131r. Anchor kutular\u0131, \u00f6zellik haritas\u0131n\u0131n her bir konumunda (pikselinde) \u00f6nceden tan\u0131mlanm\u0131\u015f, farkl\u0131 boyut ve en-boy oranlar\u0131na sahip referans kutulard\u0131r. RPN, payla\u015f\u0131lan \u00f6zellik haritas\u0131n\u0131n \u00fczerine k\u00fc\u00e7\u00fck bir kayar pencere (sliding window) uygulayarak \u00e7al\u0131\u015f\u0131r. Bu kayar pencere, bir evri\u015fimsel katman gibidir. RPN, bu iki \u00e7\u0131k\u0131\u015f\u0131 (nesne\/arka plan s\u0131n\u0131fland\u0131rmas\u0131 ve s\u0131n\u0131r kutusu kaymalar\u0131) kullanarak \u00e7ok say\u0131da aday b\u00f6lge \u00fcretir.<\/p>\n RPN, bir g\u00f6r\u00fcnt\u00fc i\u00e7in binlerce hatta on binlerce aday b\u00f6lge \u00fcretebilir. Bu b\u00f6lgelerin \u00e7o\u011fu birbiriyle \u00f6rt\u00fc\u015f\u00fcr ve ayn\u0131 nesneyi i\u015faret eder. Bu fazla ve \u00f6rt\u00fc\u015fen kutular\u0131 elemek i\u00e7in Non-Maximum Suppression (NMS) algoritmas\u0131 kullan\u0131l\u0131r. RPN taraf\u0131ndan \u00fcretilen aday b\u00f6lgelerin (ROI’ler) boyutlar\u0131 farkl\u0131l\u0131k g\u00f6sterebilir. Ancak, Fast R-CNN’in s\u0131n\u0131fland\u0131rma ve regresyon i\u00e7in kulland\u0131\u011f\u0131 tam ba\u011flant\u0131l\u0131 katmanlar, sabit boyutlu giri\u015f bekler. ROI Havuzlama katman\u0131, bu sorunu \u00e7\u00f6zmek i\u00e7in kullan\u0131l\u0131r. ROI Havuzlama katman\u0131ndan \u00e7\u0131kan sabit boyutlu \u00f6zellik vekt\u00f6rleri, a\u011f\u0131n son k\u0131sm\u0131na beslenir. Bu k\u0131s\u0131m, Fast R-CNN’deki s\u0131n\u0131fland\u0131rma ve regresyon ba\u015fl\u0131\u011f\u0131na benzer. Faster R-CNN’in e\u011fitimi, iki ana bile\u015fenin (RPN ve Fast R-CNN tespit a\u011f\u0131) birlikte optimize edilmesini gerektiren \u00e7ok g\u00f6revli bir kay\u0131p fonksiyonu kullan\u0131r. Bu, modelin hem nesne b\u00f6lgelerini do\u011fru bir \u015fekilde \u00f6nermesini hem de bu b\u00f6lgelerdeki nesneleri do\u011fru bir \u015fekilde s\u0131n\u0131fland\u0131rmas\u0131n\u0131 ve konumland\u0131rmas\u0131n\u0131 \u00f6\u011frenmesini sa\u011flar.<\/p>\n Toplam kay\u0131p fonksiyonu, RPN’in kayb\u0131 ve Fast R-CNN tespit a\u011f\u0131n\u0131n kayb\u0131n\u0131n birle\u015fimidir: Burada: Regresyon kay\u0131plar\u0131 genellikle yaln\u0131zca pozitif \u00f6rnekler i\u00e7in (yani ger\u00e7ek bir nesneye atanan anchorlar veya ROI’ler i\u00e7in) hesaplan\u0131r.<\/p>\n Faster R-CNN’in iki ana bile\u015feni (RPN ve Fast R-CNN tespit a\u011f\u0131) ayn\u0131 ortak evri\u015fimsel katmanlar\u0131 payla\u015ft\u0131\u011f\u0131 i\u00e7in, bu bile\u015fenlerin nas\u0131l e\u011fitilece\u011fi konusunda birka\u00e7 strateji geli\u015ftirilmi\u015ftir:<\/p>\n 1. 4 Ad\u0131ml\u0131 Alternatif E\u011fitim (4-Step Alternating Training):<\/strong> Bu, orijinal Faster R-CNN makalesinde \u00f6nerilen e\u011fitim y\u00f6ntemidir ve en yayg\u0131n kullan\u0131lan\u0131d\u0131r. 2. Yakla\u015f\u0131k Birlikte E\u011fitim (Approximate Joint Training):<\/strong> Bu y\u00f6ntem, 4 ad\u0131ml\u0131 e\u011fitime g\u00f6re daha basittir. RPN ve Fast R-CNN ba\u015fl\u0131klar\u0131, tek bir a\u011f i\u00e7inde birlikte e\u011fitilir. Ancak, RPN’in ileri ge\u00e7i\u015fi s\u0131ras\u0131nda \u00fcretilen b\u00f6lge \u00f6nerileri, Fast R-CNN ba\u015fl\u0131\u011f\u0131n\u0131n geri yay\u0131l\u0131m\u0131 (backpropagation) s\u0131ras\u0131nda sabit tutulur. Bu, RPN’in gradyanlar\u0131n\u0131n ortak evri\u015fimsel katmanlara geri akmas\u0131na izin vermez. Bu y\u00f6ntem, daha h\u0131zl\u0131 olmakla birlikte bazen hafif bir do\u011fruluk kayb\u0131na neden olabilir.<\/p>\n 3. Birlikte E\u011fitim (End-to-End Joint Training):<\/strong> Bu, en ideal e\u011fitim y\u00f6ntemidir. RPN ve Fast R-CNN ba\u015fl\u0131klar\u0131, ortak evri\u015fimsel katmanlar dahil olmak \u00fczere, tek bir a\u011f olarak tamamen u\u00e7tan uca e\u011fitilir. Bu, t\u00fcm a\u011f\u0131n tek bir optimize edici (\u00f6rne\u011fin SGD) ile optimize edildi\u011fi anlam\u0131na gelir ve RPN’den gelen gradyanlar ortak evri\u015fimsel katmanlara da akar. Bu, en iyi performans\u0131 sa\u011flamakla birlikte, uygulanmas\u0131 biraz daha karma\u015f\u0131k olabilir ve daha fazla bellek gerektirebilir.<\/p>\n Faster R-CNN, nesne tespiti alan\u0131nda \u00e7\u0131\u011f\u0131r a\u00e7an bir model olmas\u0131na ra\u011fmen, her teknoloji gibi kendine \u00f6zg\u00fc avantaj ve dezavantajlara sahiptir.<\/p>\n * Y\u00fcksek Do\u011fruluk:<\/strong> R-CNN ve Fast R-CNN’e k\u0131yasla, RPN’in \u00f6\u011frenilen b\u00f6lge \u00f6nerileri sayesinde daha do\u011fru tespitler yapar. Geleneksel b\u00f6lge \u00f6neri algoritmalar\u0131n\u0131n aksine, RPN nesne tespiti g\u00f6revi i\u00e7in optimize edilmi\u015ftir. * Karma\u015f\u0131kl\u0131k:<\/strong> Tek a\u015famal\u0131 tespit\u00e7ilere (YOLO, SSD) k\u0131yasla daha karma\u015f\u0131k bir mimariye ve e\u011fitim s\u00fcrecine sahiptir. \u00d6zellikle 4 ad\u0131ml\u0131 alternatif e\u011fitim, dikkatli bir uygulama gerektirir. Faster R-CNN, nesne tespiti alan\u0131nda bir d\u00f6n\u00fcm noktas\u0131 olmu\u015f ve bir\u00e7ok ger\u00e7ek d\u00fcnya uygulamas\u0131n\u0131n geli\u015ftirilmesinde temel te\u015fkil etmi\u015ftir.<\/p>\n * Otonom S\u00fcr\u00fc\u015f:<\/strong> Ara\u00e7lar\u0131n yoldaki di\u011fer ara\u00e7lar\u0131, yayalar\u0131, trafik i\u015faretlerini ve engelleri tespit etmesi i\u00e7in hayati \u00f6neme sahiptir. Faster R-CNN, bu t\u00fcr sistemlerde nesne konumland\u0131rma ve s\u0131n\u0131fland\u0131rma i\u00e7in kullan\u0131l\u0131r. Faster R-CNN, modern nesne tespiti modellerinin \u00e7o\u011funun temelini atm\u0131\u015ft\u0131r. Ancak, ara\u015ft\u0131rmalar daha h\u0131zl\u0131, daha do\u011fru ve daha verimli modeller geli\u015ftirmeye devam etmi\u015ftir. Bu yeni nesil modeller, Faster R-CNN’in sa\u011flad\u0131\u011f\u0131 temeller \u00fczerine in\u015fa edilmi\u015ftir ve s\u00fcrekli olarak performans s\u0131n\u0131rlar\u0131n\u0131 zorlamaktad\u0131r. Faster R-CNN’in mimarisi ve prensipleri, bu daha yeni modellerin anla\u015f\u0131lmas\u0131 i\u00e7in hala temel bir referans noktas\u0131d\u0131r.<\/p>\n Faster R-CNN, derin \u00f6\u011frenmenin nesne tespiti alan\u0131nda ger\u00e7ek bir devrim yaratmas\u0131n\u0131 sa\u011flayan, d\u00f6n\u00fcm noktas\u0131 niteli\u011finde bir mimaridir. B\u00f6lge \u00f6neri mekanizmas\u0131n\u0131 (RPN) derin \u00f6\u011frenme a\u011f\u0131na entegre ederek, kendisinden \u00f6nceki R-CNN ve Fast R-CNN modellerinin h\u0131z ve do\u011fruluk s\u0131n\u0131rlamalar\u0131n\u0131 a\u015fm\u0131\u015ft\u0131r. U\u00e7tan uca e\u011fitilebilir yap\u0131s\u0131, y\u00fcksek do\u011fruluk oran\u0131 ve geleneksel y\u00f6ntemlere k\u0131yasla \u00fcst\u00fcn performans\u0131 sayesinde, otonom s\u00fcr\u00fc\u015ften t\u0131bbi g\u00f6r\u00fcnt\u00fclemeye kadar pek \u00e7ok alanda nesne tespiti uygulamalar\u0131n\u0131n temelini olu\u015fturmu\u015ftur.<\/p>\n Her ne kadar g\u00fcn\u00fcm\u00fczde YOLO, SSD ve Transformer tabanl\u0131 modeller gibi daha h\u0131zl\u0131 ve bazen daha do\u011fru alternatifler ortaya \u00e7\u0131km\u0131\u015f olsa da, Faster R-CNN’in temel prensipleri ve mimari yenilikleri, nesne tespiti alan\u0131ndaki derin \u00f6\u011frenme algoritmalar\u0131n\u0131n anla\u015f\u0131lmas\u0131 i\u00e7in kritik \u00f6neme sahiptir. B\u00f6lge \u00f6neri a\u011f\u0131, anchor kutular\u0131 ve \u00e7ok g\u00f6revli kay\u0131p fonksiyonu gibi kavramlar, modern bilgisayar g\u00f6r\u00fc\u015f\u00fc ara\u015ft\u0131rmalar\u0131nda hala merkezi rol oynamaktad\u0131r. Faster R-CNN, derin \u00f6\u011frenmenin bilgisayar g\u00f6r\u00fc\u015f\u00fc alan\u0131ndaki g\u00fcc\u00fcn\u00fc ve potansiyelini a\u00e7\u0131k\u00e7a ortaya koyan, ikonik ve ilham verici bir model olarak tarihe ge\u00e7mi\u015ftir.<\/p>\n","protected":false},"excerpt":{"rendered":"Faster R-CNN: Nesne Tespitinde Devrim Yaratan Bir Mimari\nGiri\u015f: Nesne Tespitinin \u00d6nemi ve Evrimi\nNesne tespiti (object detection), bilgisayar g\u00f6","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-32412","1":"post","2":"type-post","3":"status-publish","4":"format-standard","6":"category-genel","7":"cs-entry","8":"cs-video-wrap"},"yoast_head":"\n
\n1. \u0130ki A\u015famal\u0131 (Two-Stage) Tespit\u00e7iler:<\/strong> Bu modeller, \u00f6nce aday nesne b\u00f6lgelerini (region proposals) belirler, ard\u0131ndan bu b\u00f6lgelerdeki nesneleri s\u0131n\u0131fland\u0131r\u0131r ve s\u0131n\u0131r kutular\u0131n\u0131 iyile\u015ftirir. Do\u011fruluk a\u00e7\u0131s\u0131ndan genellikle daha ba\u015far\u0131l\u0131d\u0131rlar. R-CNN, Fast R-CNN ve Faster R-CNN bu kategoriye girer.
\n2. Tek A\u015famal\u0131 (One-Stage) Tespit\u00e7iler:<\/strong> Bu modeller, aday b\u00f6lge \u00f6nerisi ad\u0131m\u0131n\u0131 atlayarak, tek bir a\u011fda hem konum hem de s\u0131n\u0131fland\u0131rma tahminlerini do\u011frudan yaparlar. Genellikle daha h\u0131zl\u0131d\u0131rlar ancak do\u011frulukta bir miktar \u00f6d\u00fcn verebilirler. YOLO (You Only Look Once) ve SSD (Single Shot MultiBox Detector) bu kategoriye \u00f6rnek olarak verilebilir.<\/p>\nR-CNN Ailesi: Evrimin Ad\u0131mlar\u0131<\/h2>\n
R-CNN (Regions with Convolutional Neural Networks): \u0130lk Ad\u0131m<\/h3>\n
\n1. B\u00f6lge \u00d6nerisi (Region Proposal):<\/strong> Se\u00e7ici Arama (Selective Search) gibi geleneksel algoritmalar kullan\u0131larak bir g\u00f6r\u00fcnt\u00fcden yakla\u015f\u0131k 2000 adet aday nesne b\u00f6lgesi (region proposals) \u00e7\u0131kar\u0131l\u0131r. Bu b\u00f6lgeler, nesne i\u00e7erme potansiyeli olan alanlard\u0131r.
\n2. \u00d6zellik \u00c7\u0131kar\u0131m\u0131:<\/strong> Her bir aday b\u00f6lgesi, sabit bir boyuta yeniden boyutland\u0131r\u0131l\u0131r ve \u00f6nceden e\u011fitilmi\u015f bir CNN (\u00f6rne\u011fin, AlexNet) arac\u0131l\u0131\u011f\u0131yla ge\u00e7irilerek derin \u00f6zellik vekt\u00f6rleri elde edilir.
\n3. S\u0131n\u0131fland\u0131rma:<\/strong> Elde edilen \u00f6zellik vekt\u00f6rleri, her bir b\u00f6lgedeki nesnenin s\u0131n\u0131f\u0131n\u0131 tahmin etmek i\u00e7in e\u011fitilmi\u015f Destek Vekt\u00f6r Makineleri (Support Vector Machines – SVM) ile s\u0131n\u0131fland\u0131r\u0131l\u0131r.
\n4. S\u0131n\u0131r Kutusu Regresyonu:<\/strong> Her s\u0131n\u0131f i\u00e7in ayr\u0131 ayr\u0131 e\u011fitilmi\u015f bir regresyon modeli, \u00f6nerilen s\u0131n\u0131r kutular\u0131n\u0131 daha do\u011fru hale getirmek i\u00e7in kullan\u0131l\u0131r.<\/p>\n
\n* Yava\u015fl\u0131k:<\/strong> Her bir aday b\u00f6lgesi i\u00e7in ayr\u0131 ayr\u0131 CNN ge\u00e7i\u015fi yap\u0131lmas\u0131, modeli a\u015f\u0131r\u0131 derecede yava\u015flat\u0131yordu (bir g\u00f6r\u00fcnt\u00fc i\u00e7in onlarca saniye).
\n* \u00c7ok A\u015famal\u0131 E\u011fitim:<\/strong> CNN, SVM’ler ve s\u0131n\u0131r kutusu regresyonu ayr\u0131 ayr\u0131 e\u011fitilmeliydi, bu da karma\u015f\u0131k bir e\u011fitim s\u00fcreci anlam\u0131na geliyordu.
\n* Disk Alan\u0131:<\/strong> Her bir b\u00f6lgenin \u00f6zellik vekt\u00f6rleri diskte depolanmal\u0131yd\u0131, bu da b\u00fcy\u00fck veri k\u00fcmeleri i\u00e7in devasa disk alan\u0131 gerektiriyordu.<\/p>\nFast R-CNN: H\u0131z ve Verimlilik Art\u0131\u015f\u0131<\/h3>\n
\nFast R-CNN’in \u00e7al\u0131\u015fma ad\u0131mlar\u0131:
\n1. B\u00f6lge \u00d6nerisi:<\/strong> Yine Se\u00e7ici Arama gibi geleneksel y\u00f6ntemlerle aday b\u00f6lgeler \u00fcretilir. Bu ad\u0131m hala bir darbo\u011fazd\u0131r.
\n2. Ortak \u00d6zellik \u00c7\u0131kar\u0131m\u0131:<\/strong> T\u00fcm g\u00f6r\u00fcnt\u00fc, tek seferde bir CNN’den ge\u00e7irilerek bir \u00f6zellik haritas\u0131 (feature map) elde edilir. Bu, R-CNN’deki en b\u00fcy\u00fck farkt\u0131r; her b\u00f6lge i\u00e7in ayr\u0131 ayr\u0131 CNN ge\u00e7i\u015fi yap\u0131lmaz.
\n3. ROI Havuzlama:<\/strong> Her bir aday b\u00f6lgesi, orijinal g\u00f6r\u00fcnt\u00fcdeki konumundan \u00f6zellik haritas\u0131 \u00fczerindeki kar\u015f\u0131l\u0131k gelen “\u0130lgi B\u00f6lgesi”ne (Region of Interest – ROI) e\u015flenir. ROI Havuzlama katman\u0131, bu farkl\u0131 boyutlardaki ROI’lerden sabit boyutlu \u00f6zellik vekt\u00f6rleri \u00e7\u0131kar\u0131r. Bu, sonraki tam ba\u011flant\u0131l\u0131 katmanlar i\u00e7in gereklidir.
\n4. S\u0131n\u0131fland\u0131rma ve Regresyon:<\/strong> Sabit boyutlu \u00f6zellik vekt\u00f6rleri, tam ba\u011flant\u0131l\u0131 katmanlara beslenir. Bu katmanlar, ayn\u0131 anda hem nesnenin s\u0131n\u0131f\u0131n\u0131 tahmin eden bir softmax s\u0131n\u0131fland\u0131r\u0131c\u0131ya hem de s\u0131n\u0131r kutusunu iyile\u015ftiren bir regresyon modeline ba\u011flan\u0131r.<\/p>\n
\n* H\u0131z:<\/strong> T\u00fcm g\u00f6r\u00fcnt\u00fc i\u00e7in tek bir CNN ge\u00e7i\u015fi sayesinde, \u00f6zellik \u00e7\u0131kar\u0131m s\u00fcreci b\u00fcy\u00fck \u00f6l\u00e7\u00fcde h\u0131zland\u0131.
\n* Tek A\u015famal\u0131 E\u011fitim (\u00c7o\u011funlukla):<\/strong> S\u0131n\u0131fland\u0131rma ve s\u0131n\u0131r kutusu regresyonu, tek bir derin \u00f6\u011frenme a\u011f\u0131 i\u00e7inde e\u011fitilebilir hale geldi. Bu, e\u011fitim s\u00fcrecini basitle\u015ftirdi ve performans\u0131 art\u0131rd\u0131.
\n* \u00c7ok G\u00f6revli Kay\u0131p Fonksiyonu:<\/strong> Hem s\u0131n\u0131fland\u0131rma hem de regresyon i\u00e7in birle\u015fik bir kay\u0131p fonksiyonu kullan\u0131ld\u0131.<\/p>\nFaster R-CNN: B\u00f6lge \u00d6neri A\u011f\u0131n\u0131n Entegrasyonu<\/h2>\n
Mimariye Genel Bak\u0131\u015f<\/h3>\n
\n1. Ortak Evri\u015fimsel \u00d6zellik \u00c7\u0131kar\u0131c\u0131 (Shared Convolutional Feature Extractor):<\/strong> G\u00f6r\u00fcnt\u00fcden derin \u00f6zellikler \u00e7\u0131karan temel bir CNN (genellikle VGG-16, ResNet gibi \u00f6nceden e\u011fitilmi\u015f bir a\u011f).
\n2. B\u00f6lge \u00d6neri A\u011f\u0131 (Region Proposal Network – RPN):<\/strong> \u00d6zellik haritas\u0131 \u00fczerinde \u00e7al\u0131\u015farak nesne i\u00e7erme potansiyeli olan aday b\u00f6lgeleri (region proposals) \u00fcreten k\u00fc\u00e7\u00fck bir sinir a\u011f\u0131.
\n3. ROI Havuzlama Katman\u0131 (ROI Pooling Layer):<\/strong> RPN taraf\u0131ndan \u00f6nerilen farkl\u0131 boyutlardaki b\u00f6lgelerden sabit boyutlu \u00f6zellik vekt\u00f6rleri \u00e7\u0131karan katman.
\n4. S\u0131n\u0131fland\u0131rma ve S\u0131n\u0131r Kutusu Regresyonu (Classification and Bounding Box Regression):<\/strong> Sabit boyutlu \u00f6zellik vekt\u00f6rlerini kullanarak nesnelerin s\u0131n\u0131flar\u0131n\u0131 tahmin eden ve s\u0131n\u0131r kutular\u0131n\u0131 iyile\u015ftiren tam ba\u011flant\u0131l\u0131 katmanlar.<\/p>\nB\u00f6lge \u00d6neri A\u011f\u0131 (RPN) Detayl\u0131 \u0130nceleme<\/h3>\n
Anchor Kutular\u0131 (Anchor Boxes)<\/h4>\n
\n* Tan\u0131m ve \u00d6nemi:<\/strong> Bir anchor kutusu, belirli bir \u00f6l\u00e7ek (scale) ve en-boy oran\u0131 (aspect ratio) ile tan\u0131mlanm\u0131\u015f bir dikd\u00f6rtgendir. \u00d6rne\u011fin, 3 farkl\u0131 \u00f6l\u00e7ek (\u00f6rne\u011fin 128×128, 256×256, 512×512 pikseller) ve 3 farkl\u0131 en-boy oran\u0131 (\u00f6rne\u011fin 1:1, 1:2, 2:1) kullan\u0131ld\u0131\u011f\u0131nda, \u00f6zellik haritas\u0131n\u0131n her bir konumunda 3×3=9 adet anchor kutusu \u00fcretilir. Bu anchor kutular\u0131, RPN’in potansiyel nesnelerin farkl\u0131 \u015fekil ve boyutlardaki varyasyonlar\u0131n\u0131 yakalamas\u0131na yard\u0131mc\u0131 olur.
\n* Pozitif\/Negatif \u00d6rnekler:<\/strong> E\u011fitim s\u0131ras\u0131nda, anchor kutular\u0131 ger\u00e7ek (ground truth) nesne kutular\u0131yla kar\u015f\u0131la\u015ft\u0131r\u0131l\u0131r. Bir anchor kutusu ile bir ger\u00e7ek kutu aras\u0131ndaki IoU (Intersection over Union) de\u011feri belirli bir e\u015fi\u011fin (\u00f6rn. 0.7) \u00fczerindeyse veya en y\u00fcksek IoU’ya sahipse “pozitif” \u00f6rnek olarak etiketlenir (yani bir nesne i\u00e7eriyor). IoU de\u011feri d\u00fc\u015f\u00fck bir e\u015fi\u011fin (\u00f6rn. 0.3) alt\u0131ndaysa “negatif” \u00f6rnek olarak etiketlenir (yani arka plan). Ara de\u011ferdeki anchor kutular\u0131 genellikle e\u011fitimde kullan\u0131lmaz.<\/p>\nRPN’in \u00c7al\u0131\u015fma Prensibi<\/h4>\n
\n1. Evri\u015fimsel Katmanlar:<\/strong> Payla\u015f\u0131lan \u00f6zellik haritas\u0131ndan gelen her bir konum i\u00e7in, k\u00fc\u00e7\u00fck bir 3×3 evri\u015fim katman\u0131 (256 veya 512 \u00e7\u0131k\u0131\u015f kanall\u0131) uygulan\u0131r. Bu katman, her bir konum i\u00e7in daha y\u00fcksek seviye \u00f6zellikler \u00e7\u0131kar\u0131r.
\n2. S\u0131n\u0131fland\u0131rma Katman\u0131 (cls layer):<\/strong> Bu \u00f6zellikler, 1×1 evri\u015fim katman\u0131na (2k \u00e7\u0131k\u0131\u015f kanall\u0131, burada k anchor say\u0131s\u0131d\u0131r) beslenir. Bu katman, her bir anchor kutusu i\u00e7in iki olas\u0131l\u0131k (nesne\/arka plan) tahmin eder. Bu, her anchor kutusunun bir nesne i\u00e7erip i\u00e7ermedi\u011fini belirler.
\n3. S\u0131n\u0131r Kutusu Regresyon Katman\u0131 (reg layer):<\/strong> Ayn\u0131 \u00f6zellikler, ba\u015fka bir 1×1 evri\u015fim katman\u0131na (4k \u00e7\u0131k\u0131\u015f kanall\u0131) beslenir. Bu katman, her bir anchor kutusu i\u00e7in 4 koordinat (x, y, geni\u015flik, y\u00fckseklik) i\u00e7eren bir kayma (offset) tahmini yapar. Bu kaymalar, anchor kutular\u0131n\u0131 ger\u00e7ek nesne kutular\u0131na daha yak\u0131n hale getirmek i\u00e7in kullan\u0131l\u0131r.<\/p>\nNon-Maximum Suppression (NMS)<\/h4>\n
\n* Ama\u00e7:<\/strong> NMS, ayn\u0131 nesneye ait birden fazla y\u00fcksek skorlu s\u0131n\u0131r kutusundan sadece en iyi olan\u0131 se\u00e7erek, nihai tahmin say\u0131s\u0131n\u0131 azalt\u0131r.
\n* \u00c7al\u0131\u015fma \u015eekli:<\/strong>
\n 1. T\u00fcm aday b\u00f6lgeler, nesne i\u00e7erme skorlar\u0131na g\u00f6re s\u0131ralan\u0131r.
\n 2. En y\u00fcksek skorlu kutu se\u00e7ilir ve nihai tahmin listesine eklenir.
\n 3. Se\u00e7ilen kutuyla belirli bir IoU e\u015fi\u011finin \u00fczerinde \u00f6rt\u00fc\u015fen di\u011fer t\u00fcm kutular listeden \u00e7\u0131kar\u0131l\u0131r.
\n 4. Kalan kutular aras\u0131ndan tekrar en y\u00fcksek skorlu kutu se\u00e7ilir ve s\u00fcre\u00e7 tekrarlan\u0131r.
\n 5. Bu i\u015flem, listede kutu kalmayana kadar devam eder.
\nNMS sonucunda, her nesne i\u00e7in tek ve en do\u011fru s\u0131n\u0131r kutusu elde edilmi\u015f olur. RPN genellikle NMS sonras\u0131 en y\u00fcksek skorlu 2000 veya 3000 b\u00f6lgeyi Fast R-CNN k\u0131sm\u0131na iletir.<\/p>\nROI Havuzlama (ROI Pooling) Katman\u0131<\/h3>\n
\n* Sabit Boyutlu \u00d6zellik Vekt\u00f6rleri:<\/strong> ROI Havuzlama, her bir ROI’yi \u00f6zellik haritas\u0131 \u00fczerinde tan\u0131mlar ve bu b\u00f6lgeden sabit boyutlu bir \u00f6zellik haritas\u0131 (\u00f6rne\u011fin, 7×7) \u00e7\u0131kar\u0131r. Bunu yaparken, ROI’yi (genellikle e\u015fit karelere) b\u00f6ler ve her bir alt b\u00f6lgeden maksimum de\u011feri (max-pooling) al\u0131r. Bu sayede, farkl\u0131 boyutlardaki ROI’lerden her zaman ayn\u0131 boyutta (\u00f6rne\u011fin, 7x7x512) bir \u00f6zellik vekt\u00f6r\u00fc elde edilir.
\n* Fast R-CNN’den Fark\u0131:<\/strong> Fast R-CNN’de ROI Havuzlama, Se\u00e7ici Arama’dan gelen b\u00f6lgeler \u00fczerinde \u00e7al\u0131\u015f\u0131rken, Faster R-CNN’de RPN’den gelen b\u00f6lgeler \u00fczerinde \u00e7al\u0131\u015f\u0131r. Temel i\u015flev ayn\u0131d\u0131r.<\/p>\nS\u0131n\u0131fland\u0131rma ve S\u0131n\u0131r Kutusu Regresyonu<\/h3>\n
\n* Tam Ba\u011flant\u0131l\u0131 Katmanlar:<\/strong> Sabit boyutlu \u00f6zellik vekt\u00f6rleri, genellikle iki veya daha fazla tam ba\u011flant\u0131l\u0131 (Fully Connected – FC) katmandan ge\u00e7er. Bu katmanlar, y\u00fcksek seviye anlamsal \u00f6zelliklerin \u00e7\u0131kar\u0131lmas\u0131na yard\u0131mc\u0131 olur.
\n* Softmax S\u0131n\u0131fland\u0131rma:<\/strong> FC katmanlar\u0131ndan sonra gelen bir katman, her bir ROI i\u00e7in nesnenin s\u0131n\u0131f olas\u0131l\u0131klar\u0131n\u0131 tahmin etmek \u00fczere bir softmax fonksiyonu kullan\u0131r. Bu, C+1<\/code> boyutlu bir vekt\u00f6r \u00e7\u0131kar\u0131r (C nesne s\u0131n\u0131f\u0131 say\u0131s\u0131, +1 arka plan s\u0131n\u0131f\u0131).
\n* S\u0131n\u0131r Kutusu Regresyonu:<\/strong> Ayn\u0131 FC katmanlar\u0131ndan sonra gelen ba\u015fka bir katman, her bir ROI i\u00e7in 4C<\/code> boyutlu bir vekt\u00f6r \u00e7\u0131kar\u0131r. Bu vekt\u00f6r, her bir s\u0131n\u0131f i\u00e7in ayr\u0131 ayr\u0131 s\u0131n\u0131r kutusu koordinatlar\u0131n\u0131 (x, y, geni\u015flik, y\u00fckseklik) iyile\u015ftirmek i\u00e7in kullan\u0131l\u0131r. Bu, modelin s\u0131n\u0131fland\u0131rma ile birlikte s\u0131n\u0131r kutular\u0131n\u0131 daha hassas hale getirmesine olanak tan\u0131r.<\/p>\nFaster R-CNN’in E\u011fitim S\u00fcreci<\/h2>\n
\u00c7ok G\u00f6revli Kay\u0131p Fonksiyonu<\/h3>\n
\n$L(p, t, p^, t^<\/em>) = L_{RPN}(p_i, t_i) + L_{FastRCNN}(p_j, t_j)$<\/p>\n
\n* $L_{RPN}$: B\u00f6lge \u00d6neri A\u011f\u0131’n\u0131n kayb\u0131d\u0131r. \u0130ki terimden olu\u015fur:
\n S\u0131n\u0131fland\u0131rma Kayb\u0131 ($L_{cls}^{RPN}$):<\/strong> Her bir anchor kutusunun nesne mi yoksa arka plan m\u0131 oldu\u011funu tahmin etmek i\u00e7in ikili \u00e7apraz entropi (binary cross-entropy) kayb\u0131 kullan\u0131l\u0131r. $p_i$ tahmin edilen olas\u0131l\u0131k, $p_i^<\/em>$ ise ger\u00e7ek etikettir (1 nesne, 0 arka plan).
\n Regresyon Kayb\u0131 ($L_{reg}^{RPN}$):<\/strong> Pozitif anchor kutular\u0131 i\u00e7in s\u0131n\u0131r kutusu kaymalar\u0131n\u0131 iyile\u015ftirmek i\u00e7in p\u00fcr\u00fczs\u00fcz L1 (smooth L1) kayb\u0131 kullan\u0131l\u0131r. $t_i$ tahmin edilen kayma, $t_i^<\/em>$ ise ger\u00e7ek kaymad\u0131r.
\n* $L_{FastRCNN}$: Fast R-CNN tespit a\u011f\u0131n\u0131n kayb\u0131d\u0131r. Bu da iki terimden olu\u015fur:
\n S\u0131n\u0131fland\u0131rma Kayb\u0131 ($L_{cls}^{FastRCNN}$):<\/strong> Her bir ROI’nin nihai s\u0131n\u0131f\u0131n\u0131 tahmin etmek i\u00e7in \u00e7apraz entropi (cross-entropy) kayb\u0131 kullan\u0131l\u0131r. $p_j$ tahmin edilen s\u0131n\u0131f olas\u0131l\u0131\u011f\u0131, $p_j^<\/em>$ ise ger\u00e7ek s\u0131n\u0131ft\u0131r.
\n Regresyon Kayb\u0131 ($L_{reg}^{FastRCNN}$):<\/strong> Her bir ROI i\u00e7in s\u0131n\u0131f baz\u0131nda s\u0131n\u0131r kutusu kaymalar\u0131n\u0131 iyile\u015ftirmek i\u00e7in p\u00fcr\u00fczs\u00fcz L1 kayb\u0131 kullan\u0131l\u0131r. $t_j$ tahmin edilen kayma, $t_j^<\/em>$ ise ger\u00e7ek kaymad\u0131r.<\/p>\nAlternatif E\u011fitim Stratejileri<\/h3>\n
\n * Ad\u0131m 1:<\/strong> RPN, ImageNet’te \u00f6nceden e\u011fitilmi\u015f bir CNN ile ba\u015flat\u0131l\u0131r ve yaln\u0131zca RPN g\u00f6revi i\u00e7in e\u011fitilir. Bu a\u015famada, tespit a\u011f\u0131 k\u0131sm\u0131 devre d\u0131\u015f\u0131d\u0131r.
\n * Ad\u0131m 2:<\/strong> RPN’den elde edilen b\u00f6lge \u00f6nerileri kullan\u0131larak, ayr\u0131 bir Fast R-CNN tespit a\u011f\u0131 e\u011fitilir. Bu a\u015famada, ortak evri\u015fimsel katmanlar ve RPN sabit tutulur.
\n * Ad\u0131m 3:<\/strong> RPN a\u011f\u0131 yeniden ba\u015flat\u0131l\u0131r, ancak bu sefer ortak evri\u015fimsel katmanlar Ad\u0131m 2’de e\u011fitilen Fast R-CNN a\u011f\u0131ndan al\u0131n\u0131r (payla\u015f\u0131lan katmanlar g\u00fcncellenir). RPN’in sadece kendine \u00f6zg\u00fc katmanlar\u0131 (s\u0131n\u0131fland\u0131rma ve regresyon) ayarlan\u0131r.
\n * Ad\u0131m 4:<\/strong> Son olarak, tespit a\u011f\u0131 yeniden e\u011fitilir. Bu sefer ortak evri\u015fimsel katmanlar ve RPN’in kendine \u00f6zg\u00fc katmanlar\u0131 da g\u00fcncellenir. Bu ad\u0131mda, RPN ve tespit a\u011f\u0131 birbiriyle uyumlu hale getirilir.
\n Bu d\u00f6ng\u00fc, modelin her iki b\u00f6l\u00fcm\u00fcn\u00fcn de ortak \u00f6zellik \u00e7\u0131kar\u0131c\u0131y\u0131 en iyi \u015fekilde kullanmas\u0131n\u0131 sa\u011flar.<\/p>\nFaster R-CNN’in Avantajlar\u0131 ve Dezavantajlar\u0131<\/h2>\n
Avantajlar\u0131<\/h3>\n
\n* H\u0131z ve Verimlilik:<\/strong> B\u00f6lge \u00f6neri a\u015famas\u0131n\u0131 da derin \u00f6\u011frenme a\u011f\u0131na entegre ederek, t\u00fcm modelin u\u00e7tan uca e\u011fitilebilir ve \u00e7ok daha h\u0131zl\u0131 \u00e7al\u0131\u015fmas\u0131n\u0131 sa\u011flar. Geleneksel y\u00f6ntemlerle kar\u015f\u0131la\u015ft\u0131r\u0131ld\u0131\u011f\u0131nda, bir g\u00f6r\u00fcnt\u00fc i\u00e7in saniyeler yerine milisaniyeler i\u00e7inde tespit yapabilir.
\n* Tek A\u011fda Entegrasyon:<\/strong> T\u00fcm bile\u015fenler (\u00f6zellik \u00e7\u0131kar\u0131c\u0131, RPN, ROI havuzlama, s\u0131n\u0131fland\u0131r\u0131c\u0131, regresyoncu) tek bir a\u011fda birle\u015ftirilmi\u015ftir. Bu, modelin y\u00f6netimini ve e\u011fitimini kolayla\u015ft\u0131r\u0131r.
\n* Esneklik:<\/strong> Temel evri\u015fimsel \u00f6zellik \u00e7\u0131kar\u0131c\u0131 (backbone) olarak farkl\u0131 CNN mimarileri (VGG, ResNet, Inception vb.) kullan\u0131labilir, bu da modelin farkl\u0131 performans\/do\u011fruluk gereksinimlerine uyarlanabilmesini sa\u011flar.
\n* Azalt\u0131lm\u0131\u015f Bellek T\u00fcketimi:<\/strong> Fast R-CNN’deki gibi \u00f6zellik haritalar\u0131n\u0131n diskte depolanmas\u0131na gerek kalmaz, bu da bellek t\u00fcketimini azalt\u0131r.<\/p>\nDezavantajlar\u0131<\/h3>\n
\n* Hesaplama Maliyeti:<\/strong> Her ne kadar R-CNN ve Fast R-CNN’e g\u00f6re \u00e7ok daha h\u0131zl\u0131 olsa da, \u00f6zellikle ger\u00e7ek zamanl\u0131 ve g\u00f6m\u00fcl\u00fc sistemler (edge devices) i\u00e7in hala y\u00fcksek hesaplama maliyeti gerektirebilir. Tek a\u015famal\u0131 modeller genellikle daha h\u0131zl\u0131d\u0131r.
\n* K\u00fc\u00e7\u00fck Nesnelerle Zorluk:<\/strong> \u00d6zellikle \u00e7ok k\u00fc\u00e7\u00fck nesnelerin tespiti konusunda hala baz\u0131 zorluklar ya\u015fayabilir. RPN’in anchor kutular\u0131 ve \u00f6zellik haritas\u0131n\u0131n \u00e7\u00f6z\u00fcn\u00fcrl\u00fc\u011f\u00fc, k\u00fc\u00e7\u00fck nesnelerin ayr\u0131nt\u0131lar\u0131n\u0131 kaybetmeye neden olabilir.
\n* Ger\u00e7ek Zamanl\u0131 Performans S\u0131n\u0131rlamalar\u0131:<\/strong> Y\u00fcksek \u00e7\u00f6z\u00fcn\u00fcrl\u00fckl\u00fc g\u00f6r\u00fcnt\u00fclerde veya k\u0131s\u0131tl\u0131 donan\u0131ma sahip sistemlerde tam ger\u00e7ek zamanl\u0131 (\u00f6rne\u011fin, 30+ FPS) performans elde etmek zor olabilir.<\/p>\nFaster R-CNN’in Uygulama Alanlar\u0131 ve Gelece\u011fi<\/h2>\n
Uygulama Alanlar\u0131<\/h3>\n
\n* G\u00fcvenlik ve G\u00f6zetim:<\/strong> G\u00fcvenlik kameralar\u0131nda \u015f\u00fcpheli nesnelerin, ki\u015filerin veya olaylar\u0131n otomatik olarak tespit edilmesi, anormal davran\u0131\u015flar\u0131n izlenmesi.
\n* T\u0131p:<\/strong> T\u0131bbi g\u00f6r\u00fcnt\u00fclerdeki (r\u00f6ntgen, MR, BT) t\u00fcm\u00f6rler, lezyonlar veya di\u011fer anormalliklerin tespiti ve te\u015fhisine yard\u0131mc\u0131 olmak.
\n* Perakende:<\/strong> Ma\u011fazalardaki \u00fcr\u00fcnlerin envanter y\u00f6netimi, m\u00fc\u015fteri davran\u0131\u015f analizi, raf d\u00fczeni optimizasyonu.
\n* End\u00fcstriyel Otomasyon:<\/strong> Fabrikalarda \u00fcretim hatt\u0131ndaki \u00fcr\u00fcnlerin kalite kontrol\u00fc, kusurlar\u0131n tespiti, robotik sistemlerin nesneleri tan\u0131mas\u0131 ve manip\u00fcle etmesi.
\n* Drone ve Robotik:<\/strong> Dronlar\u0131n ve robotlar\u0131n \u00e7evreyi alg\u0131lamas\u0131, navigasyon yapmas\u0131 ve belirli g\u00f6revleri yerine getirmesi i\u00e7in nesne tespiti temel bir yetenektir.<\/p>\nSonraki Nesil Modeller<\/h3>\n
\n* Mask R-CNN:<\/strong> Faster R-CNN’e nesne segmentasyonu (her pikselin bir nesneye ait olup olmad\u0131\u011f\u0131n\u0131 belirleme) yetene\u011fi ekleyerek, tespit edilen nesnelerin piksel d\u00fczeyinde maskelerini \u00e7\u0131karabilir.
\n* Tek A\u015famal\u0131 Modeller:<\/strong> YOLO (v3, v4, v5, v7, v8), SSD ve RetinaNet gibi modeller, h\u0131z konusunda Faster R-CNN’i geride b\u0131rakarak ger\u00e7ek zamanl\u0131 uygulamalar i\u00e7in daha uygun hale gelmi\u015ftir.
\n* Transformer Tabanl\u0131 Modeller:<\/strong> DETR (DEtection TRansformer) gibi modeller, geleneksel CNN’ler ve anchor kutular\u0131 yerine Transformer mimarisini kullanarak nesne tespiti i\u00e7in yeni bir yakla\u015f\u0131m sunmaktad\u0131r.
\n* Anchor-Free Modeller:<\/strong> FCOS, CenterNet gibi modeller, anchor kutular\u0131n\u0131n karma\u015f\u0131kl\u0131\u011f\u0131n\u0131 ortadan kald\u0131rarak nesne tespiti yapmay\u0131 hedefler.<\/p>\nSonu\u00e7<\/h2>\n