DSP-YOLO: A SAR Ship Detection Algorithm for Multiscale Sequence Fusion Based on Fusion Attention

The widespread application of synthetic aperture radar (SAR) in the domains of marine security and maritime traffic control ensures the criticality of ship detection. Nevertheless, it still faces persistent challenges due to the small size of the ship, high noise level, multiple targets, and scale variations. In this paper, we propose a fusion attention based multi-scale sequence fusion SAR ship detection algorithm called the DSP-YOLO algorithm. The proposed algorithm contains several key improvements. The C2f-DCNv2-Dynamic module, incorporating deformable convnets v2 (DCNv2) and coordinate attention (CA), is introduced to optimize the spatial and shape feature extraction of SAR ships. The neck structure of a novel Attentional Scale Sequence Fusion based YOLO(ASF-YOLO), integrated into the neck layer, optimizes the network architecture of the P2 sampling layer by reducing parameters and enhancing the extraction of raw feature information, thus improving small target detection. In addition, the Inner-PIoUv2 loss function, featuring a penalization mechanism for target size in PIoUv2 and a scaling factor in Inner-IoU, efficiently guides anchor box regression and adapts to varying target sizes. Extensive testing on the SSDD and RSDD-SAR datasets confirms the reliability and robustness of the models. The experimental results demonstrate a reduction of approximately 1/6 in the number of parameters, reaching 2.59 million compared to YOLOv8n, and the mAP50 is improved by 0.9%, reaching 99% with high accuracy on the SSDD dataset, and by 0.4% to reach 97.4% on the RSDD-SAR dataset. Compared with the existing SAR algorithms, the DSP-YOLO algorithm shows excellent performance in terms of accuracy and number of parameters.