Distribution-Based Anchor Assignment and Comprehensive Score Voting With Distance Penalty IoU Loss for SAR Remote Sensing Ship Detection
Xiaowo Xu, Xiaoling Zhang, Tianwen Zhang, Wensi Zhang, Jun Shi, Shunjun Wei, Zikang Shao, Yanqin Xu, Tianjiao Zeng
Abstract
Ship detection in synthetic aperture radar (SAR) remote sensing images is a fundamental yet challenging task in Earth observation and measurement. However, existing deep learning (DL)-based SAR ship detectors mostly struggle with unreasonable anchor assignment and limited localization quality, hindering the further accuracy increase of ship detection in SAR remote sensing images. To address the above issues, this article proposes a novel SAR ship detector, which is characterized by the distribution-based anchor assignment, comprehensive score voting, and distance penalty intersection-over-union (IoU) loss (DBAA-CSV-DPL). DBAA-CSV-DPL has three key innovations: 1) DBAA to reasonably assign anchors as positive and negative samples based on SAR scene information and anchor score distribution; 2) CSV to optimize the localization quality of detection boxes by comprehensively considering the classification and localization scores of adjacent boxes; and 3) DPL to suppress the redundant low localization quality samples by incorporating a distance penalty term. In addition, a balance-scale self-attention feature pyramid network (BS-SA-FPN) is employed to enhance the expression ability of the multiscale ship features. Extensive ablation experiments reveal each improvement’s effectiveness. Results on the SAR ship detection dataset (SSDD) and SAR-Ship-Dataset reveal the superior detection performance of DBAA-CSV-DPL, outperforming 15 state-of-the-art (SOTA) models.