SAR Image Target Recognition Based On Deep Learning Network

In recent years,the emergence and application of many deep neural networks have attracted the attention of many people.This is a new kind of signal processing algorithm,which has many advantages in the optimization of computing performance and algorithm.Naturally,these algorithms have also been used in automatic target recognition of Synthetic Aperture Radar(SAR)images,and SAR target recognition problems are similar to image classification,which uses a large number of tagged data sets to calibrate samples of unknown labels.Because of the unique imaging mechanism of SAR images,SAR images contain abundant information and are accompanied by huge speckle noise,which makes it difficult to classify and recognize the images.The advantage of the deep neural network is that it can capture more features.In this paper,SAR target recognition is based on convolution neural network,generating confrontation network and convolution generation confrontation network.The concrete work is as follows:1)We propose a parameter initialization method based on the convolution neural network.After analyzing the parameters of the traditional CNN model,We proposes a method of using the Sobel operator and the Prewitt operator in the edge filter operator in the convolution kernel to initialize the CNN network initial convolution kernel parameters.After the verification of data set of moving and stationary target acquisition and recognition(MSTAR),it is helpful to use this convolution kernel to initialize the network parameters to the convergence speed and recognition effect of the network.In practical applications,the number of training samples is often limited,so this paper proposes a supervised pre-training,using a data set for supervised training,and obtaining the network parameters to initialize the network parameters of another SAR dataset.This can also be called migration learning.This pre-training method can improve the convergence speed of the network and improve the recognition effect.2)Research on target recognition based on Generative Adversarial Networks(GAN).Considering the problem of insufficient training samples,we expand the sample data by adding different angles of conversion information.Because of the instability of the training against the network,this paper studies the effect of generating the adversarial network with matching features,and further analyzes the influence of different parameters on the recognition results.The importance and necessity of the sample label are compared and analyzed by the supervised and semisupervised target recognition experiment.3)Research on target recognition based on Deep Convolutional GAN(DCGAN).Because of the similarity of image information in spatial neighborhood,we adapt a convolutional GAN which is more suitable for extracting image features.Considering the shortage of samples,we add different angles of transformation information to experimental data.On the basis of them,we study the influence of different parameters of convolution generation on the recognition result through supervised experiments.We also analyze the quality of the generated image for the generator,and compared with the traditional shallow learning algorithm,the advantages of deep learning network are verified.