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，加权交叉熵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，数据增强与交叉熵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.
2022,44(12): 116-120 收稿日期：2022-02-25
DOI：10.3404/j.issn.1672-7649.2022.11.023
分类号：TN911.73
作者简介：冯金鹿(1987-)，男，工程师，研究方向为水声信号处理
参考文献：
[1] BU M, BENEN S, STILLER D, et al. Feature selection and classification for flase alarm redection on active diver dtection sonar data[C]//4th International Conference on Underwater Acoustics, September 2017.
[2] 关鑫, 李然威, 胡鹏. 基于数据驱动的智能化水声目标主动探测概述[J]. 船舶物资与市场, 2020(12): 1−2.
[3] 关鑫, 李然威, 胡鹏, 等. 基于改进支持向量机的水声目标-杂波不平衡分类研究[J]. 应用声学, 2021, 40(5): 715−722.
[4] Aurelien Geron. 机器学习实战: 基于Scikit-learn和Tensorflow[M]. 北京: 机械工业出版社, 2018: 30-40.
[5] 周志华. 机器学习[M]. 北京: 清华大学出版社. 2016.
[6] 张校捷. 深入浅出PyTorch从模型到源码[M]. 北京: 电子工业出版社, 2020: 163.
[7] 于化龙. 类别不平衡学习: 理论与算法[M]. 北京: 清华大学出版社, 2017: 4−5.
