SAR Image Target Recognition Based On Deep Learning

Synthetic Aperture Radar(SAR)is one type of active microwave sensors,and the target recognition techniques based on SAR images can be broadly applied in military and civilian applications.With the rapid development of deep learning theory,researchers have applied deep learning,especially deep convolutional neural network(CNN),to SAR image target recognition.When CNN is applied to SAR image target recognition tasks,how to use regularization techniques to improve the generalization ability of deep learning model is a key issue,due to the difficulty of obtaining SAR images,fewer SAR datasets and more parameters of deep CNN models.When recognizing targets with small differences,such as different subtypes in one class-variants,this problem is less studied,and target recognition accuracy is low,so optimization design of CNN model needs to be solved to improve recognition accuracy.The research has been done to solve the above problems in this thesis,and the main contents are listed as follows:Firstly,the influences of three regularization techniques on recognition accuracy,including data augmentation,dropout,and L2 regularization term,are studied using MSTAR database under standard operating conditions(SOC).The experimental results demonstrate that,the recognition accuracies of CNN models are usually improved using augmented data to perform model training.The use of dropout technique can greatly improve generalization ability of the model when AlexNet-series and ResNet-series models are used for SAR target recognition,and the influence of noises on ResNet-series models is less than other CNN models especially when dropout is used.For the models based on highway networks,The addition of L2 regularization terms can improve test accuracy of Highway-series models,but it also makes the latter phase of training extremely unstable.Secondly,a CNN architecture suitable for SAR target recognition with variants is designed,where multi-scale feature extraction strategy and the design principle of DenseNet model are adopted to preserve the input information of SAR images efficiently.The multi-scale feature extraction module is placed at the bottom of the network,where the convolution kernels with different sizes,including 1×1,3×3,5×5,7×7,9×9,are used to extract rich spatial features and meanwhile preserve the information of input image.To make information pass backwards efficiently,Dense Block module and Transition layer in DenseNet are both used to design the latter part of the model.The experimental results demonstrate that,the proposed model can obtain the classification accuracy about 95.48%when recognizing 8 subtypes of T72 target with variants.The average recognition accuracies reach 94.61%and 86.36%,respectively,when the cases of image translation and different level of noises exist in the test datasets.The recognition accuracies of 99.38%and 98.81%are achieved,respectively,under SOC,which is superior to other deep models presented in the literature.Finally,the main work and further research of the thesis is summarized.