20.500.12556/RUL-170320 Lightweight models for ocular segmentation using the IPAD method Lahki modeli za očesno segmentacijo z metodo IPAD It is known that deep neural networks are significantly overparameterized in most cases. Pruning reduces the computational complexity of the network and can even enhance its performance. To address this, we develop a framework for iterative pruning of convolutional neural networks, designed to support a variety of architectures in a user-friendly manner. We apply this framework to U-Net and SegNet models, which were trained for sclera and vein segmentation tasks on ocular images. The models were subsequently pruned using a diverse set of kernel importance measures, including the recently introduced Iterative Pruning with Activation Deviation (IPAD). We analyze the effects of pruning on model performance and compare the success of different kernel importance measures. The results we collected indicate a substantial performance advantage for pruned models over compact models trained from scratch, especially on the more challenging task of vein segmentation. We were able to confirm the superiority of the IPAD method in scenarios where ample training data is available and appropriate steps with respect to hyperparameter optimization can be taken given the computational budget. Znano je, da imajo globoke nevronske mreže v večini primerov veliko več parametrov, kot jih potrebujejo. Rezanje (angl. pruning) zmanjšuje računsko zahtevnost mreže in lahko celo izboljša njeno zmogljivost. V ta namen smo razvili orodje za iterativno rezanje konvolucijskih nevronskih mrež, ki na uporabniku prijazen način omogoča rezanje raznolikih arhitektur. Orodje smo uporabili na modelih U-Net in SegNet, naučenih za segmentacijo beločnice in žil na slikah očes. Modele smo nato rezali z uporabo različnih mer pomembnosti konvolucijskih jeder, vključno z nedavno predstavljenim pristopom Iterative Pruning with Activation Deviation (IPAD). Preučili smo vpliv rezanja na zmogljivost modelov in primerjali uspešnost različnih mer pomembnosti jeder. Zbrani rezultati kažejo na znatno prednost rezanih modelov v primerjavi z modeli, ki so bili od začetka zasnovani kot majhni, zlasti pri zahtevnejši nalogi segmentacije žil. Potrdili smo prevlado metode IPAD v primerih, ko je na voljo dovolj učnih podatkov in je izbira hiperparametrov mogoča v okviru razpoložljivega računalniškega proračuna. IPAD neural network pruning convolutional neural networks biometrics IPAD rezanje nevronskih mrež konvolucijske nevronske mreže biometrija true false false Angleški jezik Slovenski jezik Diplomsko delo/naloga 2025-07-03 14:10:03 2025-07-03 14:10:10 2025-07-18 11:22:21 0000-00-00 00:00:00 2025 0 0 0000-00-00 NiDoloceno NiDoloceno NiDoloceno 0000-00-00 0000-00-00 0000-00-00 38296 243038979 Vidovic_Matevz_-_Lahki_modeli_za_ocesno_segmentacijo_z_metodo_IPAD.pdf Vidovic_Matevz_-_Lahki_modeli_za_ocesno_segmentacijo_z_metodo_IPAD.pdf 1 4291AD88AC9B976DE795EC6179FEFE9B f568cb9376087e4a5a83ea7de99061bc73d74c49fed153ceb8e83a1560f5966f c3b4e1df-5804-11f0-b232-0050569b8976 https://repozitorij.uni-lj.si/Dokument.php?lang=slv&id=212312 Fakulteta za računalništvo in informatiko 0 0 0