Study On SAR Image Target Recognition Based On Deep Network Method

Radar automatic target recognition is mainly based on the use of one dimensional high resolution range profile or synthetic aperture radar image for the automatic classification of the targets.For the reason that the imaging mode of HRRP or SAR is complex,the features are unstable and easily affected by noise,we proposed the target recognition method for HRRP data based on sparse denoising autoencoder combined with the multi-layer perceptron deep model and the target recognition method for SAR images based on unsupervised deep networks.The models are verified by the corresponding dataset.The main work in this thesis as follows:Target recognition of HRRP data based on sparse DAE-MLP model.For the traditional HRRP target recognition methods mainly use shallow models and statistical modeling methods,and are time consuming by using the unsupevised network which needs pretrain and fine tuning process,in this thesis,in order to improve the training efficiency and use advantage of deep network in feature representation,the sparse DAE-MLP model is proposed.Firstly,the HRRP data is preprocessed,and we get the overcomplete representation of preprocessed HRRP signal by using the sparse denoising autoencoder.Lastly,the robust sparse features are used as input to train the multilayer perceptron for the target classification.The experimental results on the simulated HRRP data set show that the Sparse DAE-MLP model has a certain recognition advantage in low signal-to-noise ratio environment.Target recognition of SAR images based on the unsupervised network.Because the traditional SAR image recognition method mainly uses the method of artificial design features,such as SIFT features,linear dimensionality reduction features,in this thesis,the recognition method based on the deep belief network and the recognition method based on the stacked autoencoder are proposed.For the recognition based on DBN model,firstly,the normalization is done for SAR images,secondly,the preprocessed SAR images are used as the input of the DBN to train the model,and lastly,the testing SAR images are used to test the recognition effect of the model.For the recognition based on the stacked autoencoders,the preprocessing process of SAR image is also needed,and we get the high level features of SAR images by SAE;lastly,the SVM is used for the final classification.The experiment results on Spaceborne SAR ship data sets show that the method based on unsupervised deep network has better recognition effect than other shallow models.