YOLOv5-UOD: Underwater Object Detection Method Based on Improved YOLOv5

Xu Guan,Xiaonan Luo,Dong Wang

DOI: https://doi.org/10.1109/icist59754.2023.10367136

2023-01-01

Abstract:Underwater object detection presents a distinct and intricate challenge within the realm of computer vision. Although object detection methods have demonstrated exceptional performance on conventional datasets, the challenges associated with low visibility and color distortion in intricate underwater environments frequently lead to suboptimal image quality. Consequently, enhancing underwater object detection methods holds significant practical value. In this study, we take a pioneering approach by applying one of the most advanced object detection algorithms, YOLOv5, to underwater environments. We further tailor this approach with techniques designed specifically for underwater scenarios. Our improvements encompass the incorporation of the Swin Transformer as the foundational backbone network for YOLOv5, making the network particularly suitable for the challenges posed by underwater images featuring blurred objects. We enhance the process of feature extraction by harnessing multi-scale feature fusion, which strengthens the representation of object features in detection by integrating crossscale connections and applying context information weighting operations. This is achieved by constructing a novel feature pyramid by amalgamating the fused feature maps. Additionally, we replace the traditional Non-Maximum Suppression (NMS) with Soft-NMS, enhancing the network's capability to recognize occluded and overlapping underwater objects. The experimental outcomes provide empirical evidence supporting the efficacy of our improved network model for underwater object detection, achieving an impressive average precision (mAP) of 87.3%. Our model outperforms traditional object detection models and proves its applicability in complex underwater environments.