摘要
基于合成孔径雷达(synthetic aperture radar,SAR)在图像目标识别领域中识别精度低的问题,设计一种利用并联卷积神经网络(convolutional neural network,CNN)来提取SAR图像特征的目标识别方法.首先利用改进的ELU激活函数代替常规的ReLU激活函数,建立与二次代价函数相结合的深度学习模型.其次采用均方根支柱(root mean square Prop,RMSProp)与Nesterov动量结合的优化算法执行代价函数参数迭代更新的任务,利用Nesterov引入动量改变梯度,从两方面改进更新方式,有效地提高网络的收敛速度与精度.通过对美国国防研究规划局(DARPA)和空军研究实验室(AFRL)共同推出的MSTAR数据集进行实验,实验表明,该文提出的算法能充分提取出SAR图像中各类目标所蕴含的信息,具有较好的识别性能,是一种有效的目标识别算法.
In view of the low recognition accuracy of synthetic aperture radar(SAR)in the field of image target recognition,this paper designs a target recognition method for extracting SAR image features by shunt convolutional neural network.Firstly,the improved ELU activation function is used to replace the conventional ReLU activation function,and a deep learning model combined with the quadratic cost function is established.Secondly,this paper uses the optimization algorithm combining root mean square prop(RMSProp)and Nesterov momentum to perform the iterative updating task of cost function parameters,and uses Nesterov to introduce momentum to change the gradient to improve the updating method from two aspects to effectively improve the convergence speed and accuracy of the network.Experiments on the MSTAR data set produced by DARPA and AFRL show that the proposed algorithm is effective in target recognition since that it can extract the information contained in various targets in SAR images sufficiently,and has good recognition performance.
作者
李清
魏雪云
LI Qing;WEI Xueyun(School of Electronic Information,Jiangsu University of Science and Technology,Zhenjiang 212000,China)
出处
《电波科学学报》
EI
CSCD
北大核心
2020年第3期364-371,共8页
Chinese Journal of Radio Science
基金
国家自然科学基金(61601206)
江苏省自然科学基金(BK20160565)
江苏省高校自然科学研究项目(16KJD510001)。
