Underwater Small Object Detection in Side-Scan Sonar Images Based on Improved YOLOv5

Abstract

Automatic detection of underwater small targets is of great significance in many aspects. Aiming at the shortcomings of high missed detection and high false detection of underwater small object detection based on YOLOv5, an improved YOLOv5 method is proposed for real-time underwater small object detection in side-scan sonar images based on attention mechanism and modified anchor frames. First, the initial anchor frames of the target are re-clustered with K-means at the data input, then a new layer of feature map is added to capture the shallow features of small targets, the attention mechanism modules are introduced to extract more important features of the target, and new connections are also added to reduce the risk of overfitting caused by small samples. Experiment results show that the proposed method outperform the original and mainstream detection algorithms. The mAP value reach 96.1%, which is a 3.4% increase over the original algorithm, and the detection speed under the experimental conditions of this paper also meets the real-time requirements, which verify the effectiveness of the proposed model. Therefore, the improved YOLOv5 method can well meet the real-time requirements and provide a new strategy for underwater small object detection in side-scan sonar images.