Underwater Object Detection

This paper addresses the challenges of underwater object detection in aquaculture and marine environmental protection, particularly the issues of small targets and low clarity. To overcome these hurdles, we propose the SwinMS algorithm, which integrates the strengths of Transformer and CNN architectures to enhance the Swin Transformer. Key contributions include the introduction of a Deformable Block to improve local perception, resulting in significantly higher accuracy for underwater object detection, and the design of a Mixed-scale Block to enrich local features at each stage. Furthermore, we modify the existing activation function and enhance the detection head with the Shuffle Attention (SA) mechanism, both of which contribute to improved detection accuracy. Experimental results on the UPRC dataset demonstrate that the SwinMS algorithm achieves a mAP score of 88.6%, representing a 4.2% improvement over the DETR. These findings indicate that the proposed detection method is particularly suited for underwater object detection tasks.