SAR Ship Detection from Complex Background Based on Dynamic Shrinkage Attention Mechanism

Abstract

With the development of synthetic aperture radar (SAR) satellite, more high-resolution SAR imageries can be obtained, which have been widely used for ship detection. SAR ship detection based on convolution neural network has shown more significant potential. However, several challenges due to its imaging characteristic remain to be addressed: 1) SAR images are severely polluted by speckles under normal conditions. 2) Complex background such as inshore building and harbor may cause background to be brighter than targets. These problems result in the blurred edges of objects in SAR images and the inefficiency of ship feature extraction, which limit the performance of marine ship detection. To solve these issues, we first analyzed the impact of these issues on the feature level and proposed a feature-level denoising method called dynamic shrinkage attention (DSA). This method can adaptively suppress the irrelevant spatial response. Experiments on the SSDCB dataset show the efficiency and feasibility of our algorithm for SAR marine ship detection in complex background,