GCIE-Net: A Global and Channel Information-Enhanced Network for Ship Instance Segmentation in SAR Images

Synthetic aperture radar (SAR) ship instance segmentation is a sophisticated pixel-level analytical task that presents unique and persistent challenges in remote sensing image interpretation. In recent years, deep learning methods have attained outstanding performance and breakthroughs in SAR ship detection research field. However, there are still problems such as misjudging the background as the ships and inaccurate detection of multiscale ships. To this end, we propose a global and channel information-enhanced network (GCIE-Net), which introduces a segmentation branch into the high-performance object detection model DEIM, achieving the generation of high-quality SAR ship instance-level masks. Furthermore, we propose the multi-kernel group efficient attention mechanism (MGEA mechanism) and the cross-scale guided feature fusion module (CGFF module) to enhance the GCIE-Net from the global and channel dimensions, respectively. Specifically, the MGEA mechanism captures richer context information through dilated convolutions with different dilation rates and models global information using an efficient self-attention mechanism to improve the discrimination ability of the GCIE-Net for objects and backgrounds. The CGFF generates feature weights through cross-scale channel information to guide the feature fusion of the GCIE-Net’s encoder, achieving precise segmentation of ships at different scales. Extensive experiments conducted on the PSeg-SSDD and HRSID demonstrate the advancement and effectiveness of the proposed GCIE-Net.