Underwater Object Detection Based on Improved YOLOv8

This paper proposes the underwater target detection algorithm CMUOD based on YOLOv8 to address the issues of varying scales, low clarity, and overlapping of underwater targets. The main contribution of this article includes introducing an attention mechanism to extract key attention regions from a large amount of global information, thereby further reducing the model’s missed and false detection rates. In addition, we propose the Multi-scale Hybrid Space Pyramid Pool Fast (MHSPPF) module, which can avoid losing more information in max pooling and improve the detection performance of the model. Finally, we conducted comparative experiments on the UPR dataset and demonstrated the results. The experimental results show that the mAP score of our proposed CMUOD algorithm is 89.7 %, which is 5.3% higher than DETR. This indicates that the proposed detection method is particularly suitable for underwater object detection tasks.