In recent years,target detection in sonar images using deep learning has become a hot research topic.However,sonar images have problems such as the concentration of target scale distribution and the difficulty of data acquisition,which makes the detection effect difficult to meet the requirements.A target detection method based on variable scale prior frame is proposed to address this issue.Firstly,considering the particularity of target scale distribution in sonar images,variable scale prior frames are generated based on prior statistics.Secondly,in order to solve the problem of sonar image scarcity,data augmentation methods are used to expand the training set.Finally,the lightweighting of the model is explored by deleting the large object detection layer of the model,simplifying the model structure without reducing model accuracy.In order to evaluate the effectiveness of the algorithm,comprehensive experiments are conducted on forward-looking sonar images as an example to determine the mean average precision(mAP)@0.75 and mAP@0.5:0.95 reached 0.585 and 0.559 respectively,which increased by 5.8％and 3.1％compared to the original Yolov5 network,while giga floating-point operations(GFLOPs)decreased to 14.9.The results show that the proposed algorithm has higher accuracy and a light weight model structure.
关键词
声呐图像/目标检测/数据增强/尺度聚类/轻量化模型
Key words
sonar image/target detection/data enhancement/scale clustering/lightweight model
维普
万方数据