Research On SAR Image Recognition Based On Deep Learning

Synthetic Aperture Radar(SAR),as an active earth observation system,can be installed on flight platforms such as airplanes,satellites,and spacecraft.It can obtain remote sensing data all-time and all-weather,with high resolution and strong penetration and vitality.So it is widely used in military and civilian fields.As a unique reconnaissance method,the main application of SAR in the military field is to realize the detection and identification of specific military targets and promote the information,modernization and efficiency of military battlefields.Therefore,how to realize SAR image target recognition has important theoretical and practical significance.In recent years,the performance of computer hardware has improved rapidly,providing a good platform for the acquisition of big data,promoting the development of new artificial intelligence algorithms based on deep learning,and achieving quite good results in the field of image target detection and recognition.It has set off a research upsurge of experts and scholars at home and abroad.Due to the variability of SAR system targets,it is difficult for traditional SAR image recognition methods to fully and effectively mine the characteristics of the original data,which limits the practical application of SAR to a large extent.This paper uses the excellent feature extraction and expression capabilities of deep learning to improve the performance of SAR image target recognition,and launches related research on SAR image target recognition based on deep learning.The main research contents are summarized as follows:Aiming at the time-consuming and labor-intensive design of classifiers in SAR image target recognition and the difficulty in selecting feature extraction algorithms,in this paper,convolutional neural networks(CNNs)are introduced,and their main structures are studied in detail.The key to the improvement of CNNs network performance is summarized,and at the same time combining the theory of compressed sensing,we successfully constructed a SAR image target recognition model(CSCNNs)that can greatly reduce the parameters of the fully connected layer of the network.The algorithm model uses the Dropout mechanism and L2 regularization technology to suppress overfitting.On the basis,the fully connected layer of the network is improved,and a SAR image target classification and recognition algorithm is designed,which based on Re LU nonlinear activation function,batch normalization operation and stochastic gradient descent algorithm with momentum.The experimental results based on the MSTAR subclass data set verify the effectiveness of the model.Not only that,the experimental results of the algorithm on the ordinary natural image dataset also further verify the generality of the algorithm.Aiming at the problem of slow network convergence and over-fitting caused by the random initialization of parameters in the SAR image target recognition problem,this paper uses a supervised pre-trained convolutional neural network based on the transfer learning based on the CSCNNs network model.This method first trains the three types of military target recognition tasks under the extended conditions(EOC)of the MSTAR data set to obtain a pre-training model;then uses the pre-trained model to initialize the CSCNNs network model,and on this basis,the MSTAR data set is Ten types of military targets under standard conditions(SOC)were identified,and the classification and recognition performance is compared with the CSCNNs model with randomly initialized parameters.The final experimental results verified the effectiveness of the method.