UFL-YOLO: enhancing underwater object detection with lightweight feature refinement and noise-aware loss

Abstract

Underwater object detection is challenged by visual degradation such as low contrast, background noise and the difficulty of identifying small targets. To alleviate these issues, we propose UFL-YOLO, an enhanced version of YOLOv10 augmented with lightweight yet effective modules: small object enhance pyramid (SOEP), underwater-aware attention module (UAAM), and noise-aware focal loss (NAFL). Specifically, SOEP leverages shallow-level P2 features refined by SPDConv, fuses them with the P3 features and then processes through a cross stage partial-enclosed OmniKernel block to significantly improve the detection of small targets without adding extra detection heads. UAAM strengthens feature representation by first applying SE-style channel attention to suppress misleading activations caused by underwater color distortions, and then introducing a Sobel-guided spatial attention to enhance edge-aware localization. Furthermore, NAFL introduces an auxiliary background estimation branch to recalibrate the focal loss dynamically, thereby mitigating the influence of noisy or ambiguous regions while preserving recall for hard samples. The model is trained on URPC2020 and evaluated on UTDAC2020 without fine-tuning, achieving consistently improved detection performance and demonstrating strong cross-dataset generalization. Our implementation is publicly available at: https://github.com/wsedd/UFL-YOLO.