Research On Improved SAR Image Target Recognition Method Based On Deep Learning

Synthetic aperture radar(SAR)has working characteristics of all-weather and allweather,which has been widely adopted in security fields of military and homeland.One of the most prominent problems is synthetic aperture radar target recognition in the process of understanding and interpreting SAR images.In recent years,the research boom of artificial intelligence methods has slowly turned into the field of deep learning.Deep learning models have made remarkable research achievements in the field of image processing.Based on the imaging characteristics of SAR images,researched on SAR image target recognition method based on improved convolutional neural network.The main contents are as follows:The SAR imaging characteristics were explained.In order to solve the difficulty of using traditional target recognition methods for SAR target recognition,deep learning was introduced into SAR image target recognition methods.Firstly,the core concepts of deep learning were explained.secondly,the traditional deep learning model autoencoders and deep confidence network training methods were explained.Focused on the introduction of convolutional neural networks.By experimentally comparing the performance of convolutional neural networks with different structures,the parameters of convolutional neural network models suitable for SAR target recognition are obtained,including the number of convolutional layers and the number of convolutional kernels.Lay the foundation for improvement.Single-scale convolution kernels is adopted in ordinary convolutional neural networks,which makes other precision features easy to be ignored,so that the extracted information expression is not comprehensive.In order to solve the above problems,this paper improved the traditional convolutional neural network model and a parallel convolutional neural network model(PCNN)was proposed.The model connected convolution kernels of different sizes in parallel,processed the inputs at different scales,and then recombined them to achieve the simultaneous extraction of two scale features.In order to improve the robustness of the features,the ELU activation function was adopted.A deep learning model of combining ELU activation function and quadratic cost function was established.After part of the convolution layer,Batch Normalization(BN)was introduced to solve the problem of gradient disappearing caused by model depth and improve model training speed.Taking eight different types of SAR image data in the MSTAR data set as an example for experimental analysis,the experimental results showed that the PCNN recognition results were as high as 98.82% and have high noise resistance.Using eight types of data of different models in the same type for experimental analysis,the recognition result was as high as 95.71%,which proved the effectiveness of the method.Aiming at the problem of low accuracy in the model training process,an improved network optimization method was proposed,which was based on RMS Prop and Nesterov momentum.First,the RMS Prop algorithm was used to change the learning rate.Secondly,the Nesterov momentum was introduced to change gradient.From these two aspects,the updating method of model training parameters was improved.The improved network optimization method prevented the network from falling into a local optimal solution,thereby improving the robustness of the network model.By comparing the examples of Stochastic Gradient Descent algorithm,Ada Grad algorithm,and RMS Prop algorithm without Nesterov momentum,the effectiveness of the RMS Prop(N-RMS)optimization method with Nesterov momentum was proposed.