An Improved YOLOv8 Network for Underwater Fisheye Image Object Detection

Accurately detecting underwater objects is of great significance in fields such as marine exploration. The unique imaging principle of fisheye lenses allows for a wider field of view without compromising resolution. Although object detection technology has become relatively mature in terre-strial environments, the complex underwater environment, combined with the severe distortion of image edge regions caused by fisheye lenses, presents significant challenges for underwater object detection. In response to these issues, this paper proposed an improved YOLOv8 network architecture. A CA (Channel Attention) mechanism is added to the Bac-kbone part of YOLOv8, and a deformable convolution DCNv3 is introduced into the Bottleneck module of the C2f layer in YOLOv8. Experimental results on publicly available datasets demonstrate that the proposed method achieves higher accuracy in fisheye image object detection compared to other methods, indicating that the improvements made can enhance the feature extraction capabilities of the network structure for regions with severe fisheye image distortion. Additionally, ablation experiments are conducted to further validate the effectiveness of the proposed method.