Research Of SAR Feature Extraction And Target Recognition Based On Deep Learning

SAR image application requirements are increasing,SAR image target recognition technology is also developing.Due to the improvement of hardware performance and the introduction of effective training algorithm,the deep learning in recent years has been paid attention to and succeeded in the field of image recognition.In this thesis,the SAR feature extraction and target recognition method based on deep learning are studied by using the theory and method of deep learning and the characteristics of SAR image.The main research contents are as follows:According to the characteristics of SAR image,it points out the difficulty in SAR image target recognition.SAR image target has the characteristics of variable and uncertain,the traditional identification method requires a lot of professional knowledge,the need for image preprocessing,which can not automatically extract the effective features.Deep learning has the ability of blind learning and unsupervised learning.In this thesis,we used deep learning to solve this problem.Three kinds of deep architectures,the general neural network,deep belief network(DBN)and convolution neural network(CNN)were used for three types and ten types of SAR target recognition.By contrast,it is found that the pre-training of the deep belief network is not much improved in the ordinary neural network,and the recognition performance of the two is almost the same.The deep learning is very sensitive to the parameters and structure.It is found that the convolution neural network has great difference in SAR target recognition under different activation functions,and the Re Lu function is the most suitable as the activation function of the convolution neural network.Then,the influence of the middle structure of convolution neural network on SAR target recognition is analyzed.The pooling layer select mean-pooling and max-pooling respectively.The results are compared with the original image.The results show that the mean-pooling is more suitable as the pooling layer for SAR target recognition.Changing the convolution kernel size,we found that the most suitable convolution kernel size is releated to the size of image.This thesis also considers the recognition of SAR image with the target blocked.The recognition rate of convolution neural network was declined when the target area was half blocked.The idea of the Dropout method is that every part of the neuron is randomly discarded at each training,and the trained model has the ability to use only partial information to infer the prediction.The experimental results show that the recognition rate of neural network with Dropout is improved in the case of SAR target blocked.The comparison of the three deep architectures shows the separability of the SAR target feature extraction based on CNN is better than that of the other two,and the recognition rate of the three and ten targets is 99.8% and 96.3% respectively,which is significantly higher than that of the former two architectures.Based on the visualization of the characteristics of the convolution neural network,it is found that CNN can capture the local correlation of the SAR image and preserve the spatial structure of the image.