Erisen Serdar

Abstract

Improving the efficiency of state-of-the-art methods in semantic segmentationrequires overcoming the increasing computational cost as well as issues such asfusing semantic information from global and local contexts. Based on the recentsuccess and problems that convolutional neural networks (CNNs) encounter insemantic segmentation, this research proposes an encoder-decoder architecturewith a unique efficient residual network, Efficient-ResNet. Attention-boostinggates (AbGs) and attention-boosting modules (AbMs) are deployed by aiming tofuse the equivariant and feature-based semantic information with the equivalentsizes of the output of global context of the efficient residual network in theencoder. Respectively, the decoder network is developed with the additionalattention-fusion networks (AfNs) inspired by AbM. AfNs are designed to improvethe efficiency in the one-to-one conversion of the semantic information bydeploying additional convolution layers in the decoder part. Our network istested on the challenging CamVid and Cityscapes datasets, and the proposedmethods reveal significant improvements on the residual networks. To the bestof our knowledge, the developed network, SERNet-Former, achievesstate-of-the-art results (84.62 % mean IoU) on CamVid dataset and challengingresults (87.35 % mean IoU) on Cityscapes validation dataset.