SFL-YOLO: an improved YOLOv11-based model for underwater object detection

Object detection in complex underwater environments remains challenging due to low contrast, light scattering, noise interference, and indistinct texture and boundary information of small objects. To improve the representation capability for small objects, weak-texture objects, and objects with blurred boundaries, this study proposes an improved underwater object detection model, SFL-YOLO, based on YOLOv11n. First, a Spatial Channel Transform Convolution module (SCTC) is designed to replace conventional strided downsampling. By integrating spatial-channel rearrangement, local detail response, and detail selection gating, SCTC preserves more edge, texture, and contour information while reducing the feature map resolution. Second, a Feature-Aware Reassembly Upsampling module (FARU) is introduced. By combining content-aware reassembly, local detail guidance, and degradation-aware modulation, FARU alleviates semantic misalignment and background noise diffusion caused by fixed-interpolation upsampling. Finally, a Lightweight Detail Modeling Detection Head (LDMDH) is proposed to enhance fine-grained feature representation in the detection head while reducing parameter redundancy. Experimental results show that SFL-YOLO achieves 85.4% [email protected] on the URPC2020 dataset, improving YOLOv11n by 1.6% points. The model has 4.50 M parameters and 11.4G FLOPs, with a complete inference speed of 142 FPS. On the RUOD dataset, SFL-YOLO achieves 85.7% [email protected], outperforming YOLOv11n by 1.1% points. Ablation experiments and controlled degradation experiments further verify the effectiveness of each module and the robustness of the model on degraded underwater images. These results demonstrate that SFL-YOLO improves object detection accuracy in complex underwater environments while maintaining real-time performance.