摘要
针对合成孔径雷达(Synthetic Aperture Radar,SAR)图像目标识别算法识别准确率极易受到斑点噪声影响,且模型容易出现过拟合的问题,提出了一种基于Frost滤波和改进的卷积神经网络的目标识别方法。利用Frost滤波算法对SAR图像进行了滤波处理,减少了相干斑噪声;对CNN网络进行改进,建立了一个SAR图像目标识别模型,在网络中引入L2正则化和Dropout结构,抑制过拟合现象的发生;采用Adam优化算法,提高模型的收敛效率;最后采用组合的数据增强方法,扩充SAR图像数据集,进一步提高识别的准确率。利用美国运动和静止目标获取与识别(Moving and Stationary Target Acquisition and Recognition,MSTAR)SAR图像数据进行实验,综合识别准确率可以达到98.06%,结果表明本文所提的方法具有更好的识别效果。
Aiming at the problems that the recognition accuracy of the algorithm for recognizing Synthetic Aperture Radar(SAR)image target is easily affected by speckle noise,and the model is prone to overfitting,a novel method is proposed based on Frost filter and improved convolutional neural network.Firstly,the frost filtering algorithm was used to preprocess the SAR images to reduce the speckle noise.Secondly,some measures were taken to improve the CNN to avoid overfitting,such as establishing a model for recognizing SAR image targets and introducing L2 regulari-zation and dropout structure in the CNN.Then,the Adam optimization algorithm was used to improve the convergence efficiency of the model.Finally,the data enhancement method combining horizontal flipping,rotation,shearing and scaling was adopted to expand the SAR target data set,which can provide more sufficient samples for network train-ing,and further improve the recognition accuracy.Experiments were carried out based on the SAR image data of A-merican Moving and Stationary Target Acquisition and Recognition(MSTAR)and the comprehensive recognition ac-curacy rate can reach 98.06%.The results show that the method proposed in this paper has better recognition effect than previous researches.
作者
廉小亲
黄雪
高超
罗志宏
LIAN Xiao-qin;HUANG Xue;GAO Chao;LUO Zhi-hong(School of Artificial Intelligence,Beijing Technology and Business University,Beijing,100048,China)
出处
《计算机仿真》
北大核心
2023年第5期49-55,233,共8页
Computer Simulation
基金
北京市自然科学基金资助项目(6214034)。
关键词
合成孔径雷达
目标识别
卷积神经网络
相干斑噪声
数据增强
Synthetic aperture radar(SAR)
Target recognition(TR)
Convolutional neural network(CNN)
Speckle noise
Data augmentation
