SAR Image Target Recognition Based On Deep Learning

SAR image target recognition is to extract features from the target SAR image and determine the category attributes of the target.It has a wide range of military and civil values,such as battlefield monitoring,guided attacks,impact assessment,marine resources detection,environmental landscape detection and natural disaster assessment.The thesis focuses on deep learning methods for SAR image target recognition.The main contents include:1.Based on the characteristics of SAR images,the SAR image preprocessing methods are studied,including: filter denoising,power conversion enhancement,and energy normalization.Using the MSTAR SAR image data,the denoising performances of the median filter,adaptive median filter,Lee filter,enhanced Lee filter and Frost filter are analyzed and compared.The simulation results show that compared with other filtering methods.The enhanced Lee method can retain the details of the image better while suppressing noise.2.The SAR image recognition based on Stacked AutoEncoder(SAE)is researched.Because the SAE recognition effect is more sensitive to the structure,the thesis determines the optimal number of SAE layers and neurons by designing simulation experiments.Then using this structure to simulate the SAR image recognition effect of common stacked autoencoder,sparse stacked autoencoder and denoising stacked autoencoder,the results show that the sparse weight value and adding a certain proportion of random noise to the image can improve the model's generalization to improve recognition.To settle the problem of SAE recognition rate instability,the method combining principal component analysis(PCA)and SAE is studied.This method uses PCA to extract the main features of SAR images,then these features are identified by SAE.Simulation results show that this method can stabilize the recognition rate and accelerate the training speed compared with SAR image recognition directly using SAE.On this basis,the recognition with fusion features of PCA extraction features and image data is studied.Simulation results show that this method can improve the recognition effect.3.The SAR target recognition method based on Convolutional Neural Network(CNN)is studied.On the basis of determining the structure of the CNN,the thesis designs experiment to analyze the impact of batch size and learning rate on the recognition effect.In the selection of the activation function,the performance of the sigmoid and ReLU activation functions is analyzed through simulation experiments.The results show that the network convergence is slow and the gradient diffusion is easy to occur with the sigmoid function,and the ReLU function is better.For overfitting problem,the dropout and L2 regularization methods are introduced in CNN.The simulation results show that both methods can reduce the overfitting of the model and improve the recognition rate.Finally,the identification method combining CNN and Two-Dimensional Principle Component Analysis(2DPCA)is researched.This method uses CNN to obtain the feature maps of image data.Then these feature maps are identified after 2DPCA dimensionality reduction.The simulation results show that the proposed method is superior to the CNN and 2DPCA for the recognition of SAR images with more noise.