DRSQ-Net: underwater object detection algorithm using differentiable physical restoration and dynamic sparse query

Abstract To address detection challenges caused by color distortion, multi-scale objects, and occlusion in underwater images, this paper proposes DRSQ-Net, an improved You Only Look Once v11 (YOLOv11)-based network to enhance detection accuracy and generalization. DRSQ-Net integrates three distinctive innovations: 1) a Retinex-based differentiable physical restoration module that decomposes and optimizes underwater images to boost input feature quality; 2) a local–global collaborative Transformer enhancement network using dynamic convolution and an improved visual state-space model to capture fine details and long-range dependencies; and 3) a dynamic sparse query detection head that decouples classification and regression via task-aware routing for more robust object detection. Experiments on RUOD, URPC2020, and UTDAC2020 show that DRSQ-Net achieves 92.1% [email protected] with only 4.2 M parameters, 5.5% higher than YOLOv11. This lightweight model is well-suited for resource-constrained underwater embedded platforms, offering good practical value.