基于TGAF特征和卷积神经网络的雷达一维距离像识别

HRRP target recognition based on triple gramian angular field with convolutional neural network

基于深度学习方法的雷达一维距离像(HRRP)目标识别大多采取将二维卷积神经网络(CNN)结构转换为一维特征提取器的方法。针对低信噪比下的一维距离像目标识别,提出一种基于TGAF-CNN的雷达HRRP目标识别算法。和传统思路不同,TGAF方法将一维距离像转化为二维图像。和GAF方法相比,TGAF方法通过组合调整过的时序信息实现特征融合,提高了识别精度和鲁棒性。输出的TGAF特征图作为二维卷积神经网络的输入进行目标分类识别。基于枪械模型数据的实验结果表明,TGAF-CNN算法相对传统的深度学习方法提高了约4%的识别率。

High resolution range profile(HRRP)target recognition based on deep learning methods is mainly dedicated to changing the 2-Dimensional(2-D)convolutional neural network(CNN)framework into a 1-Dimensional(1-D)feature extractor.A new algorithm called triple gramian angular field with CNN(TGAF-CNN)is proposed for HRRP recognition in the low signal-to-noise(SNR)condition.Different from traditional methods,HRRP is transformed into a 2-D image by TGAF.Compared with GAF feature,TGAF feature combine adjusted temporal information for higher recognition precision and better robustness.TGAF feature is used as the input of 2-D CNN for recognition.Experiments based on real data from gun models indicate that TGAF-CNN can improve the HRRP recognition accuracy by about 4%compared to conventional deep learning methods.