Edge-Constrained Guided Feature Perception Network for Ship Detection in SAR Images

The ship target detection technology in synthetic aperture radar (SAR) imaging plays an important role in information warfare. However, due to the coherence of the system and the scattering characteristics of the target, SAR images are prone to speckle noise pollution, which seriously affects detection performance. To address these issues, we propose an edge-constrained guided feature perception network (ECFP-Net). Unlike the general detector, our method fully analyzes the characteristics of noise and the target’s structure, enabling the network to suppress speckle noise and extract both the local and global discrete structure features of the target. First, in the preprocessing stage, we propose a filtering method based on modified kernel (FMK), which retains the complete image structure and effectively suppresses speckle noise. Then, we design an edge constraint structure (ECS) to separate the foreground and background of the ship target, enhancing the expression ability of the ship’s structural features and assisting in positioning. Finally, we use the semantic feature perception module (SFPM) to adaptively obtain local features and semantic information by combining the vision transformer (VIT) and gating selection mechanisms. The experiment shows that the proposed ECFP-Net can achieve the most advanced level on the SAR ship detection dataset (SSDD), with average precision (AP) and F1 scores of 97.8% and 96.1%, respectively.