SAR Images Recognition Based On Convolutional Neural Network

Synthetic Aperture Radar(SAR)images target recognition has been widely used in military reconnaissance and geo-mapping,and the Automatic Target Recognition(ATR)of SAR images is an important research subject in this field.In this paper,the Convolutional Neural Network(CNN)is applied to the ground target recognition.The research is mainly about the model building and parameters optimization.The technologies involved in the construction of the model,the adaptive changes of learning rate,the optimum connection between the different layers and modify the objective function and so on.Firstly,the research background and status of SAR image target recognition are described,and CNN is studied with the development and application.The paper make an analogy between the biological neural network and the artificial neural network.The connection mode of the network is analyzed with the example of the full-connected layers.The paper introduces the common methods of data preprocessing and weight initialization.Then the convolutional layer,the pooling layer,the full-connected layer and the Softmax classifier are studied in the CNN.Finally,the backpropagation algorithm is deduced in detail.A method of SAR image ATR based on CNN is given.The paper analyzes the influence of different network depths,convolutional kernel sizes and convolutional kernel numbers on the network recognition ability through experiments,and selects the suitable model parameters for SAR image recognition.After that,adaptive learning rate algorithm is used to improve the convergence speed.An improved method named multi-scale CNN is proposed.By using asymmetric and multi-size convolutional kernels in adjacent convolutional layers to improve the recognition ability of network.And the method enriches the diversity of the extraction features of the convolutional layer.The Dropout algorithm solves the over-fitting problem in the convergence process.The paper introduces the categorical crossentropy to improve the backpropagation algorithm.By comparing with the classical methods and single scale CNNs,the experiments verifies the applicability and superiority of the method.Lastly,the main work of the paper is summarized.Then the last chapter analyzes the problems in the research,and points out the research direction in the future.