Weakly Supervised Semantic Segmentation of PolSAR Image Based on Improved SEAM

Abstract

Weakly supervised semantic segmentation (WSSS) has been widely studied in optical image field. Self-supervised equivariant attention mechanism (SEAM) effectively improves the WSSS results with the image-level labels. However, when it is directly used in the WSSS of polarimetric synthetic aperture radar (PolSAR) image, the performance is very poor. In this paper, an improved SEAM (ISEAM) is proposed for WSSS of PolSAR image, which uses the improved ResNet as the backbone network. The improvement mainly includes two aspects. First, the structure of ResNet is lightweight, which aims to match the characteristics of PolSAR dataset. Second, a squeeze-and-extraction (SE) attention mechanism is added into the backbone network, which aims to obtain channel-wise information. Experiments on Flevoland and San Francisco datasets show that the proposed ISEAM can achieve better performance than the original SEAM.