A Small-Sample Target Detection Method of Side-Scan Sonar Based on CycleGAN and Improved YOLOv8

Because of their low cost and ease of deployment, side-scan sonars is one of the most widely used underwater survey instruments. However, the complexity of the marine environment and the difficulty in target acquisition limit the detection accuracy of side-scan sonars. To address these issues, this study proposes a small-sample target detection method of side-scan sonar images using the Cycle-Consistent Generative Adversarial Network (CycleGAN) model and the improved YOLOv8 model. First, considering the difficulty in obtaining side-scan sonar target images, the proposed method uses the CycleGAN model to generate pseudo-side-scan sonar images from optical images for data augmentation. Second, the original YOLOv8 model is revised by adding attention mechanisms, employing deformable convolution networks, and updating the loss function to improve the target detection accuracy of side-scan sonar images. Experimental results demonstrated the effectiveness of the CycleGAN model in generating pseudo-side-scan sonar images and better performance of the improved YOLOv8 model in the target detection of side-scan images. Moreover, the combination of data augmentation and the improved YOLOv8 model can remarkably increase the target detection accuracy of side-scan images. The proposed methods can effectively improve the target detection efficiency of underwater sonars in marine surveying.