YOLOv13-Based Robust Underwater Object Detection via Illumination-Invariant Deformable Alignment

Abstract

To address the problems of false detections, missed detections, and unstable localization in underwater target detection under non-uniform illumination, turbidity-induced scattering, and occlusion, this paper proposes a robust underwater object detection method based on YOLOv13. The proposed method is designed around two key bottlenecks: discriminative reliability under non-uniform illumination and localization stability under occlusion and overlap. First, a learnable Illumination Invariant Module (IIM) is constructed based on illumination-invariant color-ratio theory to suppress brightness-field interference through zero-mean constrained convolution and feature fusion. Second, Bi-level Routing Attention (BRA) is incorporated into the high-level feature-fusion stage to selectively aggregate reliable contextual information and reduce interference from pseudo-textures and background highlights. Third, Deformable Convolution (DCN) is introduced into the key prediction layers to improve localization robustness in occluded and overlapping scenes through adaptive sampling and structural alignment. Experimental results show that the proposed method achieves an mAP of 95.24% at 115 FPS on the UT dataset and 84.91% at 113 FPS on the URPC2020 dataset, outperforming the YOLOv13 baseline by 2.31 and 0.76 percentage points, respectively, while maintaining favorable real-time performance. Ablation studies further verify the effectiveness of the coordinated integration of IIM, BRA, and DCN across the feature representation, feature-fusion, and prediction stages. These results demonstrate that the proposed method can provide more stable and accurate underwater target detection under complex imaging conditions.