基于卷积神经网络的主动声呐图像不平衡分类研究  

Research on unbalanced active sonar images classification based on convolutional neural network

作　　者：冯金鹿 胡鹏 何荣钦 FENG Jin-lu;HU Peng;HE Rong-qin(The 715 Research Institute of CSSC,Hangzhou 310023,China)

机构地区：[1]中国船舶集团有限公司第七一五研究所,浙江杭州310023

出　　处：《舰船科学技术》2022年第12期116-120,共5页Ship Science and Technology

摘　　要：主动声呐目标面临“大数据,小样本”的问题,“大数据”是指回波亮点数量、种类、回波形态分布非常丰富,“小样本”是指人类感兴趣的目标亮点稀少。在实际水声样本中存在无法确认标签的问题,只得将关注类型的样本标记为正类,非关注的样本标记为负类,导致负类样本远远高于正类样本,加剧了主动声呐图像集的不平衡性。为解决不平衡数据集对卷积神经网络模型的影响,本文从数据与算法2个层面研究了不平衡数据集处理方法。通过大量的实验结果表明,加权交叉炳损失函数法为处理极度不平衡主动声纳图像数据集的较好方法,并且本文两两组合数据层面与算法层面的非平衡数据集处理技术,数据增强与交叉爛损失函数相结合的模型得到具有最优的分类结果。Active sonar targets are faced the problem of big data and small samples.Big data refers to the number‘classes and morphologies of echo highlights are very rich,while small samples means that the targets highlights which human beings interested are rare.In the actual underwater acoustic samples,there is a problem of not being able to confirm the labels,so we have to mark the concerned samples as positive class,and the unconcerned samples as negative class,leading to the extreme imbalance of the active sonar image set.In order to solve the influence of unbalance data set on convolutional neural network model,this paper studies the processing methods of unbalanced data set from two levels of data and algorithm.A large number of experimental results show that the weighted cross entropy loss function method is a better method to deal with extremely unbalanced sonar data sets.In addition,this paper combines the unbalanced data set processing technologies and the data level and the algorithm level in pairs and obtains the optimal classification result by combining data enhancement with weighted cross entropy loss function.

关 键 词：卷积神经网络 不平衡数据集 主动声纳图像 特征分类 

分 类 号：TN911.73[电子电信—通信与信息系统]