Group Equivariant U-Net for the Semantic Segmentation of SAR Images

Abstract

Semantic segmentation is one of the most important steps in the analysis and interpretation of SAR images. Recently, deep segmentation networks have attracted attention because of their powerful feature extraction and classification ability. However, SAR image semantic segmentation is still a challenging task due to the lack of labeled samples for these networks. In this study, group equivariant U-Net is proposed to overcome this problem. It is aimed with the proposed method to obtain a natural generalization of U-Net without needing a data augmentation process by exploiting larger groups of symmetries, including rotation and reflections. The experiments conducted on real SAR images show that superior semantic segmentation results have been obtained with the proposed method compared to widely used alternatives.