Side‐Scan Sonar Image‐Based Object Detection With YOLOv8

The detection methods of side‐scan sonar images based on YOLOv3 and YOLOv5 suffer from issues such as insufficient feature extraction capability, high miss rate for small targets and limited localisation accuracy. In order to solve these problems, this paper discusses a side‐scan sonar detection method based on YOLOv8. YOLOv8s exploits the C2f module, decoupled detection head, anchor‐free design and dynamic label assignment strategy. This gives it stronger feature extraction capability in low‐contrast high‐noise sonar images. To study how dataset size affects model performance, overfitting and generalisation, two side‐scan sonar datasets are established in this paper. One is a small dataset with 664 images, and the other is an expanded large dataset containing 2838 images. Both datasets cover three types of underwater targets: shipwreck, person and plane. The experimental results are as follows. On the small dataset, YOLOv8s achieves the best overall performance with an mAP@0.5 of 85.5%. It outperforms YOLOv5s by 5.9% and YOLOv3‐tiny by 26.0%. Its detection accuracy improves greatly for small targets such as person and plane. After dataset expansion, all three models gain higher accuracy and overfitting is effectively reduced. YOLOv3‐tiny has the largest accuracy growth, it has a relative mAP@0.5 increase of 40.2%, whereas YOLOv8s only rises by 1.9%. This shows that YOLOv8s has strong small‐sample learning ability and good resistance to overfitting. All three models meet the real‐time detection requirements of underwater engineering. Although YOLOv8s has slightly more parameters and computational cost, it delivers the best localisation accuracy and target confidence and it works well for blurry and fragmented underwater targets. In conclusion, considering detection accuracy, generalisation, anti‐overfitting capacity and real‐time inference speed, YOLOv8s is more suitable for complex underwater target detection based on side‐scan sonar images. The research results can provide technical references for underwater target recognition scenarios including maritime search and rescue, underwater archaeology and marine engineering inspection.