基于残差网络ResNet18-SVM的弹道中段目标识别
Midcourse Ballistic Target Recognition Based on ResNet18-SVM
杨书涵 1韦楠楠 1张兴敢1
作者信息
- 1. 南京大学电子科学与工程学院,江苏南京 210023
- 折叠
摘要
针对单纯基于深度学习的弹道目标识别算法在小样本下鲁棒性差、识别率低的问题,提出一种残差网络ResNet18 与支持向量机(SVM)级联的分类模型,对弹头、重诱饵、轻诱饵和碎片四类典型弹道目标进行识别.该识别以雷达回波信号的时频图像作为输入,利用残差单元自动抽象出目标特征,再输入SVM识别分类,该模型结合了ResNet深层感受野大、SVM对高维特征样本拥有较好的分类能力的优势.文中实验采用的雷达回波数据集通过弹道仿真和目标电磁仿真获得,通过仿真实验表明级联模型ResNet18-SVM比单纯的ResNet18 识别率平均提升 1.9%;文中模型具有良好的鲁棒性,在不同信噪比下相比于SqueezeNet、ZFNet、AlexNet网络平均精准率分别高出 14.3%、16.3%、4.46%.
Abstract
In view of the problems of poor robustness and low recognition rate of midcourse ballistic target recognition algorithm based solely on deep learning in small samples,a classification model combining ResNet18 and support vector machine(SVM)was proposed to identify four typical midcourse ballistic targets,including warhead,heavy decoy,light decoy and debris.The time-frequency images of the radar echo signal were used as the input of the residual network,and the target features were automat-ically extracted by the residual unit and imported to the SVM for classification.The recognition model combines the advantages of ResNet's deep receptive field and SVM's good classification ability for high-dimensional feature samples.The experimental radar echo data is obtained by STK ballistic simulation and FEKO target electromagnetic simulation.The comparison experiments based on the narrow-band radar echo signal of the target in the middle of the ballistic trajectory show that the SVM of the combined model improves the recognition rate by 1.9％on average based on the original ResNet18;Under different Signal Noise Ratios,the pro-posed model has the highest recognition rate;Also,the average precision of proposed model is 14.3％higher than SqueezeNet,16.3％higher than ZFNet,and 4.46％higher than AlexNet,which demonstrates the advantages of cascaded residual network mod-el and SVM model.
关键词
弹道目标识别/时频图/深度学习/残差网络/特征提取Key words
ballistic target recognition/time-frequency diagram/deep learning/residual network/feature extraction引用本文复制引用
出版年
2024
国家科技期刊平台
NSTL
万方数据