20.500.12556/DKUM-87741 Primerjava modernih konvolucijskih nevronskih mrež na problemu segmentiranja slik diplomsko delo Comparison of modern convolutional neural networks for image segmentation V diplomskem delu smo preučevali in analizirali rezultate arhitektur konvolucijskih nevronskih mrež na problemu binarne segmentacije. V teoretičnem delu smo preučili arhitekture konvolucijskih nevronskih mrež SegFormer, DeepLapV3+, Gated-SCNN, Segmenter, FastFCN in TopFormer. V praktičnem delu diplomskega dela smo mreže SegFormer, Segmenter, FastFCN in TopFormer učili segmentirati slike na podatkovnih množicah Cityscapes in ADE20K. Mreže smo učili binarne segmentacije vozil. Mrežo SegFormer smo še dodatno učili na problemu binarnega segmentiranja ljudi. Dobljene rezultate smo analizirali in jih ovrednotili z evalvacijskimi metrikami. Za ocenjevanje uspešnosti smo uporabili metrike točnost, priklic, natančnost, IoU in F1 oceno. Najboljše rezultate za problem binarnega segmentiranja vozil smo dobili s pomočjo mreže SegFormer na podatkovni zbirki Cityscapes. Na podatkovni zbirki ADE20K smo za enak problem segmentacije dobili najboljše rezultate za mrežo Segmenter. Najslabše sta se izkazali mreža TopFormer na podatkovni zbirki Cityscapes in mreža FastFCN na podatkovni zbirki ADE20K. In the thesis we studied and analysed the results of convolutional neural network architectures on the problem of binary segmentation. In the theoretical part we studied the architectures of convolutional neural networks SegFormer, DeepLapV3+, Gated-SCNN, Segmenter, FastFCN, and TopFormer. In the practical part of the thesis we trained the SegFormer, Segmenter, FastFCN and TopFormer networks to segment images on the Cityscapes and ADE20K datasets. We trained the network for binary segmentation. Additionally we trained the SegFormer network on the problem of binary human segmentation. The obtained results were analysed and evaluated using evaluation metrics. We used the metrics accuracy, recall, precision, IoU and F1 score to evaluate performance. The best results for the binary vehicle segmentation problem were obtained using the SegFormer network on the Cityscapes database. On the ADE20K database we obtained the best results for the Segmenter network for the same segmentation problem. The TopFormer network performed the worst on the Cityscapes dataset, while the FastFCN network performed the worst on the ADE20K dataset. binarna segmentacija konvolucijske nevronske mreže primerjava mrež evalvacijske metrike računalniški vid binary segmentation convolutional neural networks network comparison evaluation metrics computer vision true true false [I. Dukarić] Slovenski jezik Angleški jezik Diplomsko delo/naloga 2024-03-27 07:01:24 2024-04-26 09:54:40 2024-05-14 03:18:31 0000-00-00 00:00:00 2024 0 Maribor 0 Maribor 1 spletni vir (1 datoteka PDF (XI, 52 f.)) 0000-00-00 NiDoloceno NiDoloceno NiDoloceno 0000-00-00 0000-00-00 0000-00-00 004.032.26:004.932(043.2) 195250691 UN_Dukaric_Ivana_2024.pdf UN_Dukaric_Ivana_2024.pdf 1 F36F3EB22353DA2A6EF38EFC78EEA364 c663274275e85cbdf3a86cdf4bad1b9ced6a751e5f63610b4cce026216ca9950 f636759a-ec00-11ee-a30c-00155d000105 https://dk.um.si/Dokument.php?lang=slv&id=177438 Fakulteta za elektrotehniko, računalništvo in informatiko 0 0 0