G\u00f6r\u00fcnt\u00fc i\u015fleme ve bilgisayar g\u00f6r\u00fc\u015f\u00fc alanlar\u0131, son y\u0131llarda derin \u00f6\u011frenme tekniklerinin geli\u015fimiyle birlikte devrimsel nitelikte ilerlemeler kaydetmi\u015ftir. Bu ilerlemelerin en \u00e7arp\u0131c\u0131 \u00f6rneklerinden biri de nesne alg\u0131lama ve segmentasyonudur. Nesne alg\u0131lama, bir g\u00f6r\u00fcnt\u00fcdeki nesnelerin konumlar\u0131n\u0131 (s\u0131n\u0131r kutular\u0131 ile) ve s\u0131n\u0131flar\u0131n\u0131 belirlemeyi ama\u00e7larken, segmentasyon bu nesnelerin piksel d\u00fczeyinde kesin s\u0131n\u0131rlar\u0131n\u0131 \u00e7\u0131karmay\u0131 hedefler. Bu makalede, her iki g\u00f6revi tek bir modelde birle\u015ftiren ve y\u00fcksek do\u011fruluk oranlar\u0131 sunan g\u00fc\u00e7l\u00fc bir derin \u00f6\u011frenme mimarisi olan Mask R-CNN’i, TensorFlow 1.14 ve Keras \u00e7er\u00e7evesinde detayl\u0131 bir \u015fekilde inceleyece\u011fiz.<\/p>\n
Bilgisayar g\u00f6r\u00fc\u015f\u00fc, makinelerin g\u00f6rsel d\u00fcnyay\u0131 insanlar gibi “g\u00f6rmesini” ve anlamland\u0131rmas\u0131n\u0131 sa\u011flamay\u0131 ama\u00e7layan bir yapay zeka dal\u0131d\u0131r. Bu alandaki en temel ve kritik g\u00f6revlerden biri, g\u00f6r\u00fcnt\u00fclerdeki veya videolardaki belirli nesnelerin varl\u0131\u011f\u0131n\u0131, konumunu ve t\u00fcr\u00fcn\u00fc tespit etmektir. Bu s\u00fcrece “nesne alg\u0131lama” (object detection) denir. Geleneksel nesne alg\u0131lama y\u00f6ntemleri genellikle el yap\u0131m\u0131 \u00f6zellik \u00e7\u0131kar\u0131m\u0131na ve makine \u00f6\u011frenimi s\u0131n\u0131fland\u0131r\u0131c\u0131lar\u0131na dayan\u0131rken, derin \u00f6\u011frenmenin y\u00fckseli\u015fiyle birlikte bu alanda b\u00fcy\u00fck bir paradigma de\u011fi\u015fimi ya\u015fanm\u0131\u015ft\u0131r.<\/p>\n
Nesne alg\u0131lama, bir g\u00f6r\u00fcnt\u00fcdeki birden fazla nesneyi tan\u0131mlayan ve her bir nesne i\u00e7in bir s\u0131n\u0131r kutusu (bounding box) ve bir s\u0131n\u0131f etiketi atayan bir bilgisayar g\u00f6r\u00fc\u015f\u00fc g\u00f6revidir. \u00d6rne\u011fin, bir foto\u011fraftaki t\u00fcm arabalar\u0131, yayalar\u0131 ve trafik i\u015faretlerini tespit etmek, nesne alg\u0131laman\u0131n bir uygulamas\u0131d\u0131r. Bu, otonom ara\u00e7lar, g\u00fcvenlik sistemleri ve t\u0131bbi g\u00f6r\u00fcnt\u00fcleme gibi bir\u00e7ok alanda kritik bir yetenektir.<\/p>\n
Mask R-CNN (Mask Region-based Convolutional Neural Network), nesne alg\u0131lama alan\u0131nda \u00e7\u0131\u011f\u0131r a\u00e7an Faster R-CNN mimarisinin bir uzant\u0131s\u0131d\u0131r. Faster R-CNN, nesnelerin s\u0131n\u0131r kutular\u0131n\u0131 ve s\u0131n\u0131flar\u0131n\u0131 y\u00fcksek do\u011frulukla tahmin edebilirken, Mask R-CNN bu yetene\u011fe ek olarak her tespit edilen nesne i\u00e7in piksel d\u00fczeyinde bir segmentasyon maskesi olu\u015fturma kabiliyeti sunar. Bu, Mask R-CNN’i “instance segmentation” (\u00f6rnek segmentasyonu) ad\u0131 verilen bir g\u00f6revi yerine getirebilen ilk ve en etkili modellerden biri yapm\u0131\u015ft\u0131r. \u00d6rnek segmentasyonu, ayn\u0131 s\u0131n\u0131ftaki farkl\u0131 nesneleri (\u00f6rne\u011fin, bir g\u00f6r\u00fcnt\u00fcdeki her bir farkl\u0131 insan\u0131) ayr\u0131 ayr\u0131 segmentlere ay\u0131rabilme yetene\u011fidir. Mask R-CNN, bu \u00e7oklu g\u00f6revi (s\u0131n\u0131fland\u0131rma, s\u0131n\u0131r kutusu regresyonu ve maske tahmini) tek bir entegre mimaride ba\u015far\u0131yla birle\u015ftirerek, bilgisayar g\u00f6r\u00fc\u015f\u00fc uygulamalar\u0131nda yeni ufuklar a\u00e7m\u0131\u015ft\u0131r.<\/p>\n
Bu makalenin temel amac\u0131, Mask R-CNN mimarisini derinlemesine a\u00e7\u0131klamak, temel bile\u015fenlerini ve \u00e7al\u0131\u015fma prensibini detayland\u0131rmak ve \u00f6zellikle TensorFlow 1.14 ile Keras kullanarak bu modelin nas\u0131l uygulanabilece\u011fine dair pratik bilgiler sunmakt\u0131r. Okuyucular\u0131n Mask R-CNN’in teorik altyap\u0131s\u0131n\u0131 anlamalar\u0131n\u0131n yan\u0131 s\u0131ra, kendi projelerinde bu g\u00fc\u00e7l\u00fc arac\u0131 kullanabilmeleri i\u00e7in gerekli ad\u0131mlar\u0131 ve ipu\u00e7lar\u0131n\u0131 sa\u011flamak hedeflenmektedir. TensorFlow 1.14’\u00fcn se\u00e7ilmesi, bu modelin pop\u00fclerle\u015fti\u011fi ve bir\u00e7ok mevcut projenin temelini olu\u015fturdu\u011fu d\u00f6nemi yans\u0131tmas\u0131 a\u00e7\u0131s\u0131ndan \u00f6nemlidir, ancak temel prensipler daha yeni versiyonlara da uyarlanabilir.<\/p>\n
Derin \u00f6\u011frenme \u00f6ncesi d\u00f6nemlerde nesne alg\u0131lama, genellikle el yap\u0131m\u0131 \u00f6zellik \u00e7\u0131kar\u0131c\u0131lar (\u00f6rne\u011fin SIFT, HOG) ve geleneksel makine \u00f6\u011frenimi s\u0131n\u0131fland\u0131r\u0131c\u0131lar\u0131 (\u00f6rne\u011fin SVM) kullan\u0131larak ger\u00e7ekle\u015ftirilirdi. Ancak bu y\u00f6ntemler, farkl\u0131 ayd\u0131nlatma ko\u015fullar\u0131, \u00f6l\u00e7ekler ve nesne pozisyonlar\u0131 kar\u015f\u0131s\u0131nda s\u0131n\u0131rl\u0131 performans sergiliyordu.<\/p>\n
Evri\u015fimsel Sinir A\u011flar\u0131 (Convolutional Neural Networks – CNN’ler) ile birlikte nesne alg\u0131lama alan\u0131nda b\u00fcy\u00fck bir at\u0131l\u0131m ya\u015fand\u0131. CNN’ler, g\u00f6r\u00fcnt\u00fcdeki hiyerar\u015fik \u00f6zellikleri otomatik olarak \u00f6\u011frenme yetenekleri sayesinde, geleneksel y\u00f6ntemlerin s\u0131n\u0131rlamalar\u0131n\u0131 a\u015fmay\u0131 ba\u015fard\u0131. Derin \u00f6\u011frenme tabanl\u0131 nesne alg\u0131lama modelleri genel olarak iki ana kategoriye ayr\u0131l\u0131r: tek a\u015famal\u0131 (one-stage) ve iki a\u015famal\u0131 (two-stage) dedekt\u00f6rler.<\/p>\n
Mask R-CNN’in temelini anlamak i\u00e7in, R-CNN ailesinin geli\u015fimini k\u0131saca g\u00f6zden ge\u00e7irmek faydal\u0131d\u0131r:<\/p>\n
Mask R-CNN, Faster R-CNN’in \u00fczerine in\u015fa edilmi\u015f bir mimaridir. Faster R-CNN’in s\u0131n\u0131fland\u0131rma ve s\u0131n\u0131r kutusu regresyonu g\u00f6revlerine ek olarak, her bir nesne i\u00e7in piksel d\u00fczeyinde bir segmentasyon maskesi tahmin etme yetene\u011fini ekler. Bunu, \u00e7ok g\u00f6revli bir kay\u0131p fonksiyonu ve yenilik\u00e7i bir ROI Align katman\u0131 kullanarak ba\u015far\u0131r.<\/p>\n
Mask R-CNN mimarisi, a\u015fa\u011f\u0131daki ana bile\u015fenlerden olu\u015fur:<\/p>\n
Mask R-CNN, giri\u015f g\u00f6r\u00fcnt\u00fcs\u00fcnden zengin \u00f6zellikler \u00e7\u0131karmak i\u00e7in g\u00fc\u00e7l\u00fc bir evri\u015fimsel omurga a\u011f\u0131 kullan\u0131r. Genellikle ResNet (Residual Network) mimarisi kullan\u0131l\u0131r ve bu a\u011flar genellikle ImageNet gibi b\u00fcy\u00fck veri k\u00fcmelerinde \u00f6nceden e\u011fitilmi\u015f a\u011f\u0131rl\u0131klarla ba\u015flat\u0131l\u0131r. ResNet’in derinli\u011fi (\u00f6rne\u011fin ResNet-50, ResNet-101) modelin performans\u0131n\u0131 ve hesaplama maliyetini etkiler.<\/p>\n
Omurga a\u011f\u0131 taraf\u0131ndan \u00e7\u0131kar\u0131lan \u00f6zellik haritalar\u0131, Faster R-CNN’den miras kalan RPN’e beslenir. RPN’in g\u00f6revi, bir g\u00f6r\u00fcnt\u00fcdeki nesne i\u00e7erebilecek potansiyel b\u00f6lgeleri (s\u0131n\u0131r kutular\u0131 olarak) \u00f6nermektir. Bu, nesne alg\u0131lama s\u00fcrecinin ilk a\u015famas\u0131d\u0131r.<\/p>\n
RPN, bu tahminleri kullanarak y\u00fczlerce veya binlerce potansiyel nesne \u00f6nerisi \u00fcretir. Bu \u00f6neriler daha sonra NMS (Non-Maximum Suppression) ile filtrelenerek en iyi \u00f6neriler se\u00e7ilir.\n <\/li>\n<\/ul>\n
Faster R-CNN’deki ROI Pooling katman\u0131, b\u00f6lge \u00f6nerilerini (farkl\u0131 boyutlarda olabilirler) sabit boyutlu bir \u00f6zellik haritas\u0131na d\u00f6n\u00fc\u015ft\u00fcrmek i\u00e7in kullan\u0131l\u0131yordu. Ancak ROI Pooling, ondal\u0131k olmayan piksel koordinatlar\u0131n\u0131 en yak\u0131n tam say\u0131ya yuvarlad\u0131\u011f\u0131 i\u00e7in k\u00fc\u00e7\u00fck mekansal kay\u0131plara neden oluyordu. Bu kay\u0131plar, \u00f6zellikle piksel d\u00fczeyinde hassasiyet gerektiren maske segmentasyonu g\u00f6revinde kritik \u00f6neme sahiptir.<\/p>\n
ROI Align katman\u0131ndan \u00e7\u0131kan sabit boyutlu \u00f6zellik haritalar\u0131, Mask R-CNN’in “\u00e7oklu g\u00f6rev kafas\u0131”na beslenir. Bu kafa, \u00fc\u00e7 paralel \u00e7\u0131kt\u0131 dal\u0131na sahiptir ve her biri farkl\u0131 bir g\u00f6revi yerine getirir:<\/p>\n
Mask R-CNN’in \u00e7al\u0131\u015fma prensibini ad\u0131m ad\u0131m \u00f6zetleyelim:<\/p>\n
E\u011fitim (Training):<\/strong> Mask R-CNN, u\u00e7tan uca e\u011fitilir. Kay\u0131p fonksiyonu (loss function), \u00fc\u00e7 g\u00f6revin (s\u0131n\u0131fland\u0131rma kayb\u0131, s\u0131n\u0131r kutusu regresyon kayb\u0131 ve maske kayb\u0131) toplam\u0131ndan olu\u015fur. Bu \u00fc\u00e7 kay\u0131p, modelin her bir g\u00f6revi ayn\u0131 anda optimize etmesini sa\u011flar. Maske kayb\u0131, yaln\u0131zca ger\u00e7ek s\u0131n\u0131f\u0131n maske tahmini \u00fczerinde hesaplan\u0131r. Genellikle \u00f6nceden e\u011fitilmi\u015f a\u011f\u0131rl\u0131klarla (\u00f6rne\u011fin ImageNet \u00fczerinde) ba\u015flan\u0131r ve hedef veri k\u00fcmesi \u00fczerinde ince ayar (fine-tuning) yap\u0131l\u0131r.<\/p>\n \u00c7\u0131kar\u0131m (Inference):<\/strong> E\u011fitimli bir Mask R-CNN modeli, yeni bir g\u00f6r\u00fcnt\u00fcye uyguland\u0131\u011f\u0131nda yukar\u0131daki ad\u0131mlar\u0131 takip eder ve her bir alg\u0131lanan nesne i\u00e7in s\u0131n\u0131f, s\u0131n\u0131r kutusu ve segmentasyon maskesini tahmin eder. \u00c7\u0131kar\u0131m s\u0131ras\u0131nda, NMS ve g\u00fcven e\u015fikleri gibi post-processing ad\u0131mlar\u0131, nihai sonu\u00e7lar\u0131n kalitesini art\u0131rmak i\u00e7in kullan\u0131l\u0131r.<\/p>\n Mask R-CNN’in Keras ve TensorFlow 1.14 ile uygulanmas\u0131, genellikle Matterport taraf\u0131ndan sa\u011flanan a\u00e7\u0131k kaynakl\u0131 Keras Mask R-CNN implementasyonu kullan\u0131larak yap\u0131l\u0131r. Bu implementasyon, COCO veri k\u00fcmesi \u00fczerinde \u00f6nceden e\u011fitilmi\u015f a\u011f\u0131rl\u0131klar\u0131 da i\u00e7erir ve transfer \u00f6\u011frenimi i\u00e7in m\u00fckemmel bir ba\u015flang\u0131\u00e7 noktas\u0131d\u0131r.<\/p>\n Uygulamaya ba\u015flamadan \u00f6nce a\u015fa\u011f\u0131daki k\u00fct\u00fcphanelerin y\u00fckl\u00fc oldu\u011fundan emin olunmal\u0131d\u0131r:<\/p>\n GPU Deste\u011fi:<\/strong> Mask R-CNN gibi derin ve hesaplama a\u00e7\u0131s\u0131ndan yo\u011fun modellerin e\u011fitimi i\u00e7in GPU (NVIDIA CUDA uyumlu) kullan\u0131m\u0131 neredeyse zorunludur. TensorFlow’un GPU versiyonunun do\u011fru \u015fekilde kuruldu\u011fundan ve CUDA Toolkit ile cuDNN’in uyumlu s\u00fcr\u00fcmlerinin sistemde mevcut oldu\u011fundan emin olunmal\u0131d\u0131r.<\/p>\n Kendi \u00f6zel veri k\u00fcmenizi kullanacaksan\u0131z, veri haz\u0131rl\u0131\u011f\u0131 kritik bir ad\u0131md\u0131r:<\/p>\n Matterport’un Keras Mask R-CNN uygulamas\u0131, modelin yap\u0131land\u0131rmas\u0131n\u0131 ve mimarisini i\u00e7eren Kendi veri k\u00fcmeniz \u00fczerinde Mask R-CNN’i e\u011fitmek i\u00e7in, \u00f6ncelikle E\u011fitimden sonra, model yeni g\u00f6r\u00fcnt\u00fcler \u00fczerinde nesneleri alg\u0131lamak ve segmentlere ay\u0131rmak i\u00e7in kullan\u0131labilir.<\/p>\n Nesne alg\u0131lama ve segmentasyon modellerinin performans\u0131n\u0131 de\u011ferlendirmek i\u00e7in \u00e7e\u015fitli metrikler kullan\u0131l\u0131r:<\/p>\n Mask R-CNN’in benzersiz yetenekleri, onu \u00e7ok \u00e7e\u015fitli pratik uygulamalar i\u00e7in ideal bir se\u00e7im haline getirmi\u015ftir:<\/p>\n Mask R-CNN, \u00f6rnek segmentasyon alan\u0131nda bir d\u00f6n\u00fcm noktas\u0131 olmas\u0131na ra\u011fmen, ara\u015ft\u0131rmalar daha h\u0131zl\u0131, daha hafif ve daha do\u011fru modeller geli\u015ftirmeye devam etmektedir.<\/p>\n Mask R-CNN, derin \u00f6\u011frenme tabanl\u0131 nesne alg\u0131lama ve \u00f6rnek segmentasyon alan\u0131nda g\u00fc\u00e7l\u00fc ve \u00e7ok y\u00f6nl\u00fc bir mimaridir. Faster R-CNN’in sa\u011flam temelleri \u00fczerine in\u015fa edilen bu model, s\u0131n\u0131fland\u0131rma, s\u0131n\u0131r kutusu regresyonu ve piksel d\u00fczeyinde maske tahmini g\u00f6revlerini tek bir a\u011fda birle\u015ftirerek bilgisayar g\u00f6r\u00fc\u015f\u00fc yeteneklerini \u00f6nemli \u00f6l\u00e7\u00fcde ileriye ta\u015f\u0131m\u0131\u015ft\u0131r. \u00d6zellikle TensorFlow 1.14 ve Keras gibi yayg\u0131n kullan\u0131lan \u00e7er\u00e7evelerle entegrasyonu, geli\u015ftiricilerin bu g\u00fc\u00e7l\u00fc modeli kendi uygulamalar\u0131na kolayca dahil etmelerini sa\u011flam\u0131\u015ft\u0131r.<\/p>\n Y\u00fcksek do\u011fruluk ve detayl\u0131 segmentasyon yetene\u011fi sayesinde Mask R-CNN, otonom s\u00fcr\u00fc\u015ften t\u0131bbi g\u00f6r\u00fcnt\u00fclemeye, roboti\u011fe kadar geni\u015f bir yelpazede kritik uygulamalarda kendine yer bulmu\u015ftur. Her ne kadar hesaplama a\u00e7\u0131s\u0131ndan yo\u011fun olsa da, transfer \u00f6\u011frenimi ve \u00f6nceden e\u011fitilmi\u015f modellerin kullan\u0131m\u0131, bu teknolojiyi daha eri\u015filebilir hale getirmektedir. Gelecekte daha h\u0131zl\u0131 ve verimli modellerin geli\u015ftirilmesiyle birlikte, Mask R-CNN’in temel prensipleri, bilgisayar g\u00f6r\u00fc\u015f\u00fcn\u00fcn bir\u00e7ok alan\u0131nda ilham kayna\u011f\u0131 olmaya devam edecektir. Bu makale, okuyuculara Mask R-CNN’in hem teorik altyap\u0131s\u0131n\u0131 hem de pratik uygulama ad\u0131mlar\u0131n\u0131 anlamalar\u0131 i\u00e7in kapsaml\u0131 bir rehber sunmay\u0131 ama\u00e7lam\u0131\u015ft\u0131r.<\/p>\n <\/body><\/p>\n","protected":false},"excerpt":{"rendered":"Mask R-CNN ile Nesne Alg\u0131lama: TensorFlow 1.14 ve Keras Kullan\u0131m\u0131\nG\u00f6r\u00fcnt\u00fc i\u015fleme ve bilgisayar g\u00f6r\u00fc\u015f\u00fc alanlar\u0131, son y\u0131llarda derin \u00f6\u011fr","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-35666","1":"post","2":"type-post","3":"status-publish","4":"format-standard","6":"category-genel","7":"cs-entry","8":"cs-video-wrap"},"yoast_head":"\nTensorFlow 1.14 ve Keras ile Mask R-CNN Uygulamas\u0131<\/h3>\n
Gerekli K\u00fct\u00fcphaneler ve Ortam Kurulumu<\/h4>\n
\n
pip install tensorflow==1.14.0<\/code><\/li>\npip install opencv-python<\/code><\/li>\npip install numpy<\/code><\/li>\npip install matplotlib<\/code><\/li>\npip install h5py<\/code><\/li>\ngit clone https:\/\/github.com\/matterport\/Mask_RCNN.git<\/code> komutuyla indirilir ve ard\u0131ndan pip install -r requirements.txt<\/code> ile ba\u011f\u0131ml\u0131l\u0131klar kurulur.<\/li>\n<\/ul>\nVeri K\u00fcmesi Haz\u0131rl\u0131\u011f\u0131<\/h4>\n
\n
utils.py<\/code> ve model.py<\/code> dosyalar\u0131ndaki veri y\u00fckleme fonksiyonlar\u0131n\u0131 kendi format\u0131n\u0131za uyarlaman\u0131z gerekebilir. VGG Image Annotator (VIA) veya LabelMe gibi ara\u00e7lar, maske anotasyonu yapmak i\u00e7in kullan\u0131labilir.\n <\/li>\nModelin Y\u00fcklenmesi ve \u00d6nceden E\u011fitilmi\u015f A\u011f\u0131rl\u0131klar<\/h4>\n
model.py<\/code> dosyas\u0131n\u0131 sa\u011flar. Genellikle, COCO veri k\u00fcmesi \u00fczerinde \u00f6nceden e\u011fitilmi\u015f a\u011f\u0131rl\u0131klarla ba\u015flamak en iyi yakla\u015f\u0131md\u0131r (transfer \u00f6\u011frenimi).<\/p>\n\nimport os\nimport sys\nimport random\nimport math\nimport numpy as np\nimport skimage.io\nimport matplotlib.pyplot as plt\n\nMatterport Mask R-CNN k\u00fct\u00fcphanesini i\u00e7e aktar<\/h2>\nROOT_DIR = os.path.abspath(\"..\/\")\nsys.path.append(ROOT_DIR)\nfrom mrcnn import utils\nimport mrcnn.model as modellib\nfrom mrcnn import visualize\n\n
Kendi yap\u0131land\u0131rma dosyan\u0131z\u0131 veya standart COCO yap\u0131land\u0131rmas\u0131n\u0131 kullan\u0131n<\/h2>\nfrom samples.coco import coco\n\n
Modelin kaydedilece\u011fi dizin<\/h2>\nMODEL_DIR = os.path.join(ROOT_DIR, \"logs\")\n\n
COCO veri k\u00fcmesi \u00fczerinde \u00f6nceden e\u011fitilmi\u015f a\u011f\u0131rl\u0131klar\u0131n yolu<\/h2>\nCOCO_MODEL_PATH = os.path.join(ROOT_DIR, \"mask_rcnn_coco.h5\")\n\n
E\u011fer a\u011f\u0131rl\u0131klar yoksa indir<\/h2>\nif not os.path.exists(COCO_MODEL_PATH):\n utils.download_trained_weights(COCO_MODEL_PATH)\n\n
Yap\u0131land\u0131rma s\u0131n\u0131f\u0131<\/h2>\nclass InferenceConfig(coco.CocoConfig):\n # Modelin \u00e7\u0131kar\u0131m modunda \u00e7al\u0131\u015fmas\u0131 i\u00e7in toplu i\u015f boyutunu (batch size) 1 olarak ayarla\n GPU_COUNT = 1\n IMAGES_PER_GPU = 1\n\nconfig = InferenceConfig()\nconfig.display()\n\n
Modeli ba\u015flat<\/h2>\nmodel = modellib.MaskRCNN(mode=\"inference\", model_dir=MODEL_DIR, config=config)\n\n
\u00d6nceden e\u011fitilmi\u015f COCO a\u011f\u0131rl\u0131klar\u0131n\u0131 y\u00fckle<\/h2>\nmodel.load_weights(COCO_MODEL_PATH, by_name=True)\n\n
S\u0131n\u0131f isimleri (COCO veri k\u00fcmesi i\u00e7in)<\/h2>\nclass_names = ['BG', 'person', 'bicycle', 'car', 'motorcycle', 'airplane',\n 'bus', 'train', 'truck', 'boat', 'traffic light',\n 'fire hydrant', 'stop sign', 'parking meter', 'bench', 'bird',\n 'cat', 'dog', 'horse', 'sheep', 'cow', 'elephant', 'bear',\n 'zebra', 'giraffe', 'backpack', 'umbrella', 'handbag', 'tie',\n 'suitcase', 'frisbee', 'skis', 'snowboard', 'sports ball',\n 'kite', 'baseball bat', 'baseball glove', 'skateboard',\n 'surfboard', 'tennis racket', 'bottle', 'wine glass', 'cup',\n 'fork', 'knife', 'spoon', 'bowl', 'banana', 'apple',\n 'sandwich', 'orange', 'broccoli', 'carrot', 'hot dog', 'pizza',\n 'donut', 'cake', 'chair', 'couch', 'potted plant', 'bed',\n 'dining table', 'toilet', 'tv', 'laptop', 'mouse', 'remote',\n 'keyboard', 'cell phone', 'microwave', 'oven',\n 'toaster', 'sink', 'refrigerator', 'book', 'clock', 'vase',\n 'scissors', 'teddy bear', 'hair drier', 'toothbrush']\n<\/code><\/pre>\n
E\u011fitim S\u00fcreci<\/h4>\n
Config<\/code> s\u0131n\u0131f\u0131ndan t\u00fcretilen kendi yap\u0131land\u0131rma s\u0131n\u0131f\u0131n\u0131z\u0131 olu\u015fturman\u0131z gerekir. Bu s\u0131n\u0131f, veri k\u00fcmesindeki s\u0131n\u0131f say\u0131s\u0131n\u0131, g\u00f6r\u00fcnt\u00fc boyutlar\u0131n\u0131, \u00f6\u011frenme oran\u0131n\u0131 ve di\u011fer hiperparametreleri tan\u0131mlar.<\/p>\n\nclass CustomConfig(coco.CocoConfig):\n # Kendi veri k\u00fcmenizin ad\u0131n\u0131 verin\n NAME = \"object\"\n \n # S\u0131n\u0131f say\u0131s\u0131n\u0131 g\u00fcncelleyin (arka plan + kendi s\u0131n\u0131flar\u0131n\u0131z)\n NUM_CLASSES = 1 + 2 # BG + 2 kendi s\u0131n\u0131f\u0131n\u0131z (\u00f6rn. \"elma\", \"muz\")\n\n # E\u011fitim i\u00e7in GPU say\u0131s\u0131 ve her GPU'daki g\u00f6r\u00fcnt\u00fc say\u0131s\u0131\n GPU_COUNT = 1\n IMAGES_PER_GPU = 2 # Daha b\u00fcy\u00fck bir de\u011fer bellek hatas\u0131na neden olabilir\n\n # Di\u011fer hiperparametreleri ayarlay\u0131n\n STEPS_PER_EPOCH = 100\n DETECTION_MIN_CONFIDENCE = 0.9\n\nconfig = CustomConfig()\nconfig.display()\n\nE\u011fitim modunda modeli ba\u015flat<\/h2>\nmodel = modellib.MaskRCNN(mode=\"training\", model_dir=MODEL_DIR, config=config)\n\n
\u00d6nceden e\u011fitilmi\u015f a\u011f\u0131rl\u0131klar\u0131 y\u00fckle (COCO veya kendi a\u011f\u0131rl\u0131klar\u0131n\u0131z)<\/h2>\n
E\u011fer kendi a\u011f\u0131rl\u0131klar\u0131n\u0131z yoksa, COCO a\u011f\u0131rl\u0131klar\u0131ndan ba\u015flan\u0131r<\/h2>\nmodel.load_weights(COCO_MODEL_PATH, by_name=True, exclude=[\"mrcnn_class_logits\", \"mrcnn_bbox_fc\", \n \"mrcnn_bbox\", \"mrcnn_mask\"])\n\n
Kendi veri k\u00fcmesi s\u0131n\u0131f\u0131n\u0131z\u0131 y\u00fckleyin (CustomDataset)<\/h2>\n
dataset_train = CustomDataset()<\/h2>\n
dataset_train.load_custom_data(dataset_dir, \"train\")<\/h2>\n
dataset_train.prepare()<\/h2>\n\n
dataset_val = CustomDataset()<\/h2>\n
dataset_val.load_custom_data(dataset_dir, \"val\")<\/h2>\n
dataset_val.prepare()<\/h2>\n\n
E\u011fitim a\u015famalar\u0131<\/h2>\n
1. A\u015fama: Yaln\u0131zca kafa katmanlar\u0131n\u0131 e\u011fit (transfer \u00f6\u011frenimi)<\/h2>\nprint(\"Kafa katmanlar\u0131n\u0131 e\u011fitiyor...\")\nmodel.train(dataset_train, dataset_val,\n learning_rate=config.LEARNING_RATE,\n epochs=10,\n layers='heads')\n\n
2. A\u015fama: T\u00fcm katmanlar\u0131 ince ayar yap<\/h2>\nprint(\"T\u00fcm katmanlara ince ayar yap\u0131yor...\")\nmodel.train(dataset_train, dataset_val,\n learning_rate=config.LEARNING_RATE \/ 10,\n epochs=20,\n layers=\"all\")\n<\/code><\/pre>\n
\n
LEARNING_RATE<\/code>, STEPS_PER_EPOCH<\/code>, EPOCHS<\/code>, DETECTION_MIN_CONFIDENCE<\/code> gibi hiperparametreler modelin performans\u0131n\u0131 ve e\u011fitim s\u00fcresini do\u011frudan etkiler. Bu de\u011ferler, veri k\u00fcmenize ve mevcut donan\u0131m\u0131n\u0131za g\u00f6re ayarlanmal\u0131d\u0131r.\n <\/li>\n\u00c7\u0131kar\u0131m (Inference)<\/h4>\n
\nTest i\u00e7in bir g\u00f6r\u00fcnt\u00fc y\u00fckle<\/h2>\n
IMAGE_DIR = os.path.join(ROOT_DIR, \"images\")<\/h2>\n
file_names = next(os.walk(IMAGE_DIR))[2]<\/h2>\n
image = skimage.io.imread(os.path.join(IMAGE_DIR, random.choice(file_names)))<\/h2>\n\n
\u00d6rne\u011fin, bir dosya yolundan resim y\u00fckle<\/h2>\nimage_path = \"path\/to\/your\/image.jpg\"\nimage = skimage.io.imread(image_path)\n\n
Modeli \u00e7\u0131kar\u0131m modunda \u00e7al\u0131\u015ft\u0131r<\/h2>\nresults = model.detect([image], verbose=1)\n\n
Sonu\u00e7lar\u0131 al<\/h2>\nr = results[0]\n\n
Sonu\u00e7lar\u0131 g\u00f6rselle\u015ftir<\/h2>\nvisualize.display_instances(image, r['rois'], r['masks'], r['class_ids'], \n class_names, r['scores'])\n<\/code><\/pre>\n
\n
model.detect()<\/code> fonksiyonu, alg\u0131lanan nesnelerin s\u0131n\u0131r kutular\u0131n\u0131 (rois<\/code>), maskelerini (masks<\/code>), s\u0131n\u0131f ID’lerini (class_ids<\/code>) ve g\u00fcven skorlar\u0131n\u0131 (scores<\/code>) i\u00e7eren bir s\u00f6zl\u00fck d\u00f6nd\u00fcr\u00fcr.\n <\/li>\nSonu\u00e7lar\u0131n De\u011ferlendirilmesi<\/h4>\n
\n
Mask R-CNN’in Avantajlar\u0131 ve Dezavantajlar\u0131<\/h3>\n
Avantajlar<\/h4>\n
\n
Dezavantajlar<\/h4>\n
\n
Kullan\u0131m Alanlar\u0131<\/h3>\n
\n
Gelecek Trendler ve Alternatifler<\/h3>\n
\n
Sonu\u00e7<\/h3>\n