Study On The Sparsity-based SAR Image Target Recognition

Synthetic Aperture Radar (SAR) is important means of earth observation. Automatic target recognition (ATR) based on SAR image is of great significance in the battlefield environment. This dissertation addresses the sparsity-based SAR ATR, especially focusing on the target recognition under occlusion and automatic feature extraction from data. The main content of this dissertation is summarized as follows.In the first part, a new classification method is proposed based on non-negative sparse representation in order to solve the occlusion issue in SAR image target recognition, The difference between LO-norm and L1-norm minimization in solving non-negative sparse representation problem is analyzed, and it is proved that L1-norm regularization method pursuits not only the sparsity of the solution but also the similarity between the input signal and the selected atoms under some conditions, hence it is fit for classification application. The experimental results on moving and Stationary Target Acquisition and Recognition (MSTAR) dataset show that the non-negative sparse representation based classification method with L1-norm regularization can achieve much better recognition performance, andit is more robust in the recognition of targets with occlusion compared with the traditional method.Although it is shown that both the sparse representation based classification and non-negative sparse representation based classification methods is robust to occlusion, they all make an assumption that the occlusion is pixel-wise, i.e. the probability of occlusion in one pixel are independent to the others. However, the effect of occlusion caused by the objects smaller than the resolution cell is similar to the speckle noise, while the effect caused objects larger than the resolution cell will emerges in a region of adjacent cells. A Structured Sparse Occlusion Model (SSOM) is proposed to characterize the test image, which tries to separate the test input into occlusion part and its sparse representation part on the training set. When these two parts are estimated, we identify the test target only by its sparse representation to avoid the impact of occlusion. The experimental results show that the proposed method achieves robustness not only to the occlusion size change but also to the shape variety and scatter fluctuation.SAR image of ground target contains the target region formed by the scattered echoes of the target as well as the shadow area. However, the characteristics of the two areas are essentially different, therefore the traditional SAR image automatic target recognition methods used mainly target area information alone or shadow region only for recognition. This section presents a joint sparse representation model by combining the shadow region and target region images. By using mixed L1\L2 norm minimization method to solve the joint sparse representation model, the SAR image target recognition is achieved by minimizing the joint reconstruction error. Recognition results on moving and stationary target acquisition and recognition (MSTAR) data sets show that the joint sparse representation model can effectively fuse the information within the target region and shadow region, and it has much better recognition performance than the methods using only the target or shadow area information of the image.In the fourth part, a new objective function for dictionary learning is presented, which adds a similarity constraint into the objective in order to enhance the discriminative in learned dictionary, while the aim in traditional dictionary learning methods, such as K-SVD, is minimize the error for well reconstruction. The new added term try to constraint the similarity between sparse representations of samples in different class be zero, i.e. the representations in feature space are differ greatly in order to be separated. The experimental results show that the similarity constrained dictionary learning perform better than other dictionary learning methods in SAR target recognition.Feature extraction is a key step in synthetic aperture radar (SAR) image target recognition. The existence of speckle and discontinuity make the conventional methods for natural images difficult to apply. Although deep belief networks (DBNs) can be used to learn feature representations automatically, they work essentially in an unsupervised way, and hence the learned features are task-irrelevant. We propose a new Boltzmann machine called similarity constrained restricted Boltzmann machines (SRBMs), which inject the supervised information into learning process through constraint on the similarity of feature vectors. Furthermore, a deep architecture named as similarity constrained deep belief networks (SDBNs) is constructed by layer-wise stacking of SRBMs. Experimental results show the proposed SDBN is superior to DBN and PCA in SAR Image target recognition.In the last part, the subjects in real world recognition system are considered, such as, invariance under target translation, invariance under speckle variation in different observations, and tolerance of pose missing in training data. The method of training data augmented with the synthesized pose images is introduced to train the classifier for target identification. Inspired by the sparse representation model, the model for synthesizing pose images is also developed, which approximately construct the missing pose image by linear combining several images available. Moreover, we investigate the capability of deep convolutional neural network (CNN) combined with three types of data augmentation operations for real situations in SAR target recognition. Experimental results demonstrate the effectiveness and efficiency of the proposed framework. Especially, the CNN trained by all the data augmentation operations shows the possibility of dealing translation, speckle noising and pose missing problems in a unify process.