DLFE-YOLO: an enhanced framework for underwater object detection based on YOLOv8

Abstract Underwater object detection remains challenging because of low visibility, complex lighting, and varied target scales, which limit the effectiveness of conventional deep learning methods. To overcome these limitations, this study proposes DLFE-You Only Look Once (YOLO), a novel YOLOv8-based framework that enhances feature representation and detection robustness in underwater environments. Its key innovations include its improved attention mechanisms and adaptive feature modeling to better capture irregular and small targets, as well as a loss function that emphasizes difficult-to-detect objects. Extensive experiments on two public benchmarks (URPC2020 and URPC2018) demonstrate that the proposed framework consistently outperforms state-of-the-art detectors, achieving mAP @0.5 values of 84.4% and 77.9%, respectively. Transfer learning experiments further indicate that DLFE-YOLO maintains strong adaptability across datasets, effectively leveraging knowledge from larger datasets to improve detection in smaller or less diverse underwater domains. These results highlight the model’s ability to address the shortcomings of existing methods. It improves both detection accuracy and generalization and provides a practical solution for intelligent underwater perception systems.