Salient Feature Pyramid Network for Ship Detection in SAR Images

Recently, convolutional neural networks (CNNs) have achieved remarkable performance in SAR ship detection. However, intense background interference in SAR images and the inherent attributes of the ship targets severely hinder further improvements in detection accuracy. To address these problems, we propose a novel anchor-free detector with a salient feature pyramid network (SFPN) and an intersection-over-union (IoU) branch. First, SFPN aggregates multi-level semantic features and generates salient features for ship targets through a saliency extraction (SE) module. Then, a saliency fusion (SF) module merges the salient features and FPN features. In particular, we introduce a heatmap branch in SFPN as an extra supervision signal to direct the network to focus on salient regions containing ship targets rather than interference. Second, considering the shape characteristic that the length of a ship target is much larger than its width, we apply the IoU branch to the detection head to predict localization confidence for each detected bounding box. Extensive experiments conducted on High-Resolution SAR Images Dataset (HRSID) and SAR-Ship-Dataset demonstrate that the proposed method effectively improves detection accuracy and achieves state-of-the-art performance.