Dual-Branch Underwater Object Detection

This paper presents enhancements to the YOLOv8 model, addressing the complexities of underwater environments and the challenges associated with achieving high accuracy in object detection tasks. The FASFFHead feature pyramid network was crafted to derive further information from underwater optical images, with the goal of mitigating the challenges associated with low-resolution imagery.A dual-branch programmable gradient network (PGI) was designed to s address the issue of information bottleneck and gradient instability, so as to boost the model's performance. In addition, the Generalized Efficient Layer Aggregation Network (GELAN) is innovatively introduced to adapt to deep learning tasks of different complexity while optimizing computational efficiency and performance. Lab observations demonstrate that the improved model achieves an <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{mAP}$</tex> of 85.0% on the URPC 2020 dataset, which not only markedly improves detection precision but also adapts to more complex tasks, demonstrating a broad application prospect.