Hierarchical Attention for Ship Detection in SAR Images

Abstract

Considering the difficulty of ship detection in Synthetic Aperture Radar (SAR) images lacking color and texture details, we propose a method for SAR ship detection based on hierarchical attention mechanism. Compared with the optical images, the detection methods based on deep-learning for SAR images are aiming at designing a network that is sensitive to high-level features. The proposed method, containing Global Attention Module (GAM) and Local Attention Module (LAM), presents a hierarchical attention strategy respectively from the image level and the target level. GAM in both spatial and channel domain is constructed to highlight target characteristics. Anchor generation is guided by the LAM for locating more accurate candidate regions. Hierarchical attention improves the performance of extracting significant features of SAR images, which makes the algorithm more efficient in detecting small and multiple ship targets. Experiments results show that our method has achieved 0.7 to 4.1 points higher Average Precision (AP) than several state-of-the-art detection methods on SAR datasets of GF-3 and Sentinel-1.