基于YOLOv3锚框优化的侧扫声呐图像目标检测

利用侧扫声呐图像来探查海底目标对海洋资源开采和海上军事防护都有重大意义。目前人为提取图像特征进行目标检测的传统机器学习方法逐渐被深度学习取代。深度学习技术在降低算法复杂度的同时提高图像目标检测效率，极大地推动了目标检测技术地发展。将深度学习检测算法应用到侧扫声呐图像目标检测领域时，锚框作为目标检测网络中较为重要的先验信息会影响最终的检测性能，考虑到声呐数据集的真实目标框与网络设定的锚框未必贴合的问题，本文在YOLOv3的基础上对锚框进行了优化，给出了一种能够获取有效先验锚框的策略。首先使用K-Means算法对真实目标框进行聚类，获得比较贴合于声呐数据集的锚框，然后设计了一种超参数锚框映射关系对聚类后的锚框进行拉伸变换，这样获得的锚框既包含了声呐数据集的目标框信息，也能利用到YOLOv3的多尺度特性。实验结果表明，所提锚框优化策略能够让YOLOv3网络获得更优的检测性能，适用于侧扫声呐图像的目标检测问题。 

Target Detection in Side Scan Sonar Images Based on YOLOv3 Anchor Boxes Optimization

? ?The use of side scan sonar images to detect submarine targets was of great significance to the exploitation of marine resources and marine military protection. At present， the traditional machine learning method of artificially extracting image features for target detection was gradually replaced by deep learning. Deep learning technology improved the efficiency of image target detection while reducing the complexity of the algorithm， which greatly promoted the development of target detection technology. When deep learning detection algorithms were applied to the field of side scan sonar image target detection， anchor boxes were important as prior information in the target detection network which affected the final detection performance. Considering the problem that real target boxes of the sonar dataset and anchor boxes set by the network may not fit， this paper optimized anchor boxes on the basis of YOLOv3， and proposed a strategy that can obtain effective prior anchor boxes. Firstly， real target boxes were clustered using the K-Means algorithm to obtain anchor boxes that fit the sonar data set， and then a hyperparameter mapping relationship was designed to stretch and transform the clustered anchor boxes. In such case， the obtained anchor boxes not only contained the target information of the sonar dataset but also took advantage of the multi-scale features of YOLOv3. The experimental results show that the proposed anchor boxes optimization strategy can make the YOLOv3 network obtain better detection performance， which is suitable for the problem of side scan sonar image target detection.