基于深度学习的主动声呐目标回波识别研究  

Active Sonar Target-Echo Recognition Research Based on Deep Learning

作　　者：潘成胜[1,2] 毛家林 杨阳[1,2] PAN Cheng-sheng;MAO Jia-lin;YANG Yang(Key Laboratory of Communication and Network,Dalian University,Dalian Liaoning 116622,China;College of Information Engineering,Dalian University,Dalian Liaoning 116622,China)

机构地区：[1]大连大学通信与网络重点实验室,辽宁大连116622 [2]大连大学信息工程学院,辽宁大连116622

出　　处：《计算机仿真》2020年第11期179-183,293,共6页Computer Simulation

基　　金：国家自然基金(61540024);辽宁省博士启动基金(20170520371)。

摘　　要：在水下小目标探测与识别研究中,目标回波受水下复杂环境干扰严重。传统的水声目标识别方法是直接对水下目标进行特征提取再利用分类器识别,但由于对目标回波特征认知有限,手工提取目标特征会不可避免的丢失一部分关键信息。针对上述问题,研究了一种基于贝叶斯正则化理论的BP(Back Propagation)神经网络识别算法。方法可以在一定程度上避免人工特征提取丢失信息的问题,提高目标识别率。依据目标亮点模型,基于贝叶斯正则化理论,推导了BP神经网络训练结果的网络性能函数,利用网络训练过程中超参数大小的自适应调整,优化网络性能,采用识别准确率和识别效率等性能参数评价效果。仿真结果表明,与传统L-M优化算法相比,基于贝叶斯正则化理论的BP神经网络算法抗干扰能力强,在优化模型的不断更新中提高了目标识别的效率和精度。In the research on underwater small target detection and recognition,target echo is seriously disturbed by complex underwater environment.The traditional method of underwater acoustic target recognition extracts underwater target features and then selects classifier to recognize them.However,some key information can be inevitably lost due to manually extracting feature.To solve these problems,a BP(Back Propagation)neural network recognition algorithm based on Bayesian regularization theory is proposed.This method can avoid the problem of losing information in artificial feature extraction to a certain extent and improve the target recognition rate.Based on the target highlight model and Bayesian regularization theory,the network performance function of BP neural network training results was derived.The network performance was optimized by adaptively adjusting the size of super-parameters in the process of network training,and the performance parameters such as recognition accuracy and recognition efficiency were used to evaluate the effect.The simulation results show that,compared with traditional L-M optimization algorithm,the BP neural network algorithm based on Bayesian regularization theory avoids the loss of features in the process of manual extraction,has strong anti-interference ability,and improves the efficiency and accuracy of target recognition through continuous updating optimization model.

关 键 词：水下目标探测 亮点模型 贝叶斯正则化 反向传播神经网络 

分 类 号：TB566[交通运输工程—水声工程]