Side-Scan Sonar Underwater Target Detection: Combining the Diffusion Model With an Improved YOLOv7 Model

Abstract

Side-scan sonar (SSS) plays a crucial role in underwater exploration. Autonomous analysis of SSS images is vital for detecting unknown targets in underwater environments. However, due to the complexity of the underwater environment, few highlighted areas of the target, blurred feature details, and the difficulty of collecting data from SSS, achieving high-precision autonomous target recognition in SSS images is challenging. This article solves this problem by improving the You Only Look Once v7 (YOLOv7) model to achieve high-precision object detection in SSS images. First, we enhance and enlarge real and experimental images using the denoising–diffusion model to establish a self-made SSS image data set, as there are data pictures of the detection target in the SSS images obtained from real experiments. Since the SSS image has large areas without targets, this article introduces a vision transformer (ViT) for dynamic attention and global modeling, which improves the model's weight in the target region. Second, the convolutional block attention module is adopted to further improve the feature expression ability and reduce floating-point operations. Finally, this article uses Scylla-Intersection over Union as the loss function to increase the accuracy of the model's inference. Experiments on the SSS image data set demonstrate that the improved YOLOv7 model outperforms other technologies, with an average accuracy (mAP0.5) and (mAP0.5:0.95) of 78.00% and 48.11%, respectively. These results are 3.47% and 2.9% higher than the YOLOv7 model. The improved YOLOv7 algorithm proposed in this article has great potential for object detection and recognition of SSS images.