Semantic segmentation using edge loss

Semantic segmentation aims to assign correct class labels per pixel on an image. In particular, semantic segmentation has difficulties in particular along the boundary of objects. Recently, ELKPPNet has been proposed, which improves the performance of semantic segmentation by adding edge loss term into the conventional semantic segmentation algorithm. It extracts edge from the end of the network, which is used in the computation of loss. In this paper, we present U-Net based networks which adopt the edge loss of the ELKPPNet. Presented algorithm computes edge by additional network flow in the U-Net. Two different network structures are investigated. One computes edge at the end of decoder in encoder-decoder of the U-Net. The other computes edge from the start of decoder in U-Net. Experimental results show that integrating edge information in semantic segmentation improves performance.