基于拉东投影与改进卷积神经网络的小样本水下目标声呐图像识别方法  

Recognition method of sonar images from a small sample underwater targets based on Radon projection and improved convolutional neural network

作　　者：周光波 张培珍 莫晴舒 尹晓锋 ZHOU Guangbo;ZHANG Peizhen;MO Qingshu;YIN Xiaofeng(College of Electronic and Information Engineering,Guangdong Ocean University,Zhanjiang 524088,China)

机构地区：[1]广东海洋大学电子与信息工程学院,广东湛江524088

出　　处：《哈尔滨工程大学学报》2024年第10期2048-2056,共9页Journal of Harbin Engineering University

基　　金：国家自然科学基金项目(11974084);广东省自然科学基金项目(2022A1515011067).

摘　　要：针对水下声呐图像质量差、样本数量少导致目标识别精确度低的问题,本文提出一种水下目标识别方法。利用增量的全向Radon投影特征图作为输入数据,结合改进结构的卷积神经网络,实现小样本声呐图像识别。实验以5种不同目标声呐图像的Radon特征图作为输入,分别采用迁移学习得到的ResNet-18、GoogLeNet模型以及改进模型进行实验,验证改进模型的结构合理性;将原始图像结合改进模型进行识别,验证Radon特征图作为数据源的优势。原图结合改进模型、Radon特征图结合ResNet-18、GoogLeNet模型及改进模型的最优训练样本数分别为960、1440、5760和1200;训练用时依次为328、699、8678和447 s;相应最佳识别准确率分别为97.8%、94.4%、93.9%和99.9%。通过混淆矩阵给出不同方法预报错误的类别及数量,进一步解释出现误判的原因。结果表明:本文所提出的方案能够在较少的样本数和较低的运算成本条件下获取较高的精度。研究成果能够作为目标声呐图像识别分类的有效方法,并可望推广至更多水下目标分类。In view of the low accuracy of target recognition caused by the poor quality and small number of samples of underwater sonar images,a new underwater target recognition method is composed.The incremental omni-directional Radon projection feature images are used as input data,combined with the convolutional neural network with the improved structure,sonar image recognition of small size sample is realized.In the experiment,five sonar images with Radon feature images are taken as input.The ResNet-18 and GoogLeNet models obtained by transfer learning and the improved models are used to verify the structural rationality of the improved models,respectively.Then,the original images are combined with the improved model for recognition to verify the advantages of Radon feature images as data source.The optimal number of training samples for the original image combined with improved model,Radon feature image combined with ResNet-18,GoogLeNet model and improved model are 960,1440,5760 and 1200,correspondingly.Furthermore,the training time are 328 s,699 s,8678 s,447 s,and the corresponding optimal recognition accuracies are 97.8%,94.4%,93.9%and 99.9%,respectively.In addition,the classification and quantity of the prediction errors of different methods are given by confusion matrix,and the causes of the misjudgments are further explained.The results show that the proposed scheme can obtain high accuracy under the conditions of fewer samples and lower time cost.The research results can be used as an effective method for target sonar image recognition and classification,and are expected to be extended to more underwater target classification.

关 键 词：水下目标识别 声呐图像 数据增量 RADON变换 卷积神经网络 迁移学习 深度学习 特征融合 

分 类 号：O427.9[理学—声学] TP181[理学—物理]