基于深度神经网络的雷达目标高分辨距离像稳健识别方法

项目名称： 基于深度神经网络的雷达目标高分辨距离像稳健识别方法

项目编号： No.61501155

项目类型： 青年科学基金项目

立项/批准年度： 2016

项目学科： 无线电电子学、电信技术

项目作者： 潘勉

作者单位： 杭州电子科技大学

项目金额： 21万元

中文摘要： 传统雷达HRRP目标识别方法应用于高速非合作目标识别将面临两大难题：训练样本数有限和用于识别的HRRP信噪比较低。此时，HRRP特征提取和分类器设计较为困难，识别性能普遍偏低。然而，这类目标威胁很大，恰是关注重点。为解决传统方法的缺陷，本课题拟从特征提取和分类器设计出发，构建基于深度神经网络的雷达自动目标识别系统。在特征提取层面，提取体现数据内在物理特性的特征，提高后续系统识别能力。为改善系统的噪声稳健性，拟采用基于先验知识的稳健玻尔兹曼机算法，该算法通过高信噪比训练样本特征内包含的先验信息对低信噪比测试样本的特征进行增强。在分类器构建层面，拟构建深度卷积神经网络进行自动目标识别。该方法将原本独立建模的各角域子模型联系到一起，共同建模学习来挖掘数据之间的联系, 减少了系统对训练样本数的需求。本课题的研究预计在小样本和低信噪比条件下能达到较好的识别性能，可提高雷达目标识别系统的工程实用性。

中文关键词： 一维距离像；特征提取；目标识别；深度卷积神经网络；稳健玻尔兹曼机

英文摘要： The heavy requirement of training sample size and low SNR of the test HRRP are two major challenges when apply traditional radar high resolution range profile (HRRP) target recognition method to high speed non-cooperative target. Under this circumstance, extracting feature and designing classifier for HRRP are very difficult. Therefore, the recognition rates of the traditional models are generally low. However, the high speed non-cooperative targets are great threats to our defending system, which should receive significant attention. To solve the defects of traditional recognition method, we developed a radar HRRP automatic recognition system based on the deep neural network which puts emphases on both feature extraction and classifier design. This recognition system extracts the physical properties of the data, which can enhance the recognition performance. In order to improve the noise robustness of the recognition system, we explore a robust Boltzmann machine based on prior knowledge, which enhances the feature of low SNR testing sample by using priori information contained in the feature of high SNR training samples. To relax the heavy requirement of training sample size, we proposed a radar HRRP automatic recognition model based on the deep convolutional neural network. Compared with the traditional recognition methods learn the model parameters with the data according to the angle independently, the deep convolutional neural network learns model parameters with the data together, which explores the connection between the data and relaxes the requirement of training sample. Our research is expected to achieve good recognition performance under the condition of small training samples and low SNR. It can greatly expand the application range of the radar HRRP automatic target recognition.

英文关键词： high resolution range profile;feature extraction;target recognition;deep convolutional neural networks;robust Boltzmann machine