A Method Based on YOLOv8 for Sonar Image Object Detection

With the rapid development of sonar technologies, sonar images have become an important basis for underwater object detection. However, due to various factors, the quality of sonar images is not good enough and sonar images generally have severe distortion, which leads to the low accuracy in sonar image detection field. Therefore, this article proposes a method using improved YOLOv8 network introducing deformable convolution operation into network's module to reduce the influence of severe distortion. At the same time, to accelerate the convergence of the network, we introduce the idea of transfer learning. Finally, we evaluate the effectiveness of the network by conducting experiments on the URPC2022 dataset. Compared with some mainstream object detection networks mentioned in this article, the improved YOLOv8 network shows the best accuracy in the indicator of mAP@0.5:0.95.