CSTransNet: A Cross-Scale Transformer Network for Underwater Object Detection

As ocean exploration and research advance, underwater object detection plays an important role and holds significant research value. However, underwater object detection faces several challenges: (1) The underwater environment is complex, typically marked by poor lighting and the abundance of plankton, which result in blurry images and color distortion. (2) The background in underwater images is intricate and highly similar to various small targets. This paper proposes the Cross-scale Transformer Network (CSTransNet) for underwater object detection. This network incorporates a Cross-scale Fusion Transformer Block designed to integrate image features of different resolutions, enabling cross-feature layer and cross-scale feature fusion, which enhances the model's ability to distinguish between complex underwater backgrounds and small targets. Additionally, the Depth Extraction Transformer Block connects pixel-level global context information and establishes long-range dependencies, thereby enhancing detection precision in challenging underwater settings, particularly for small objects. Experimental results on the DUO, AUDD, and Brackish datasets show that CSTransNet achieves 87.9%, 88.7%, and 98.6% average precision (AP50), respectively. The proposed method exhibits higher detection accuracy and holds significant potential for various underwater tasks.