SAR Target Recognition Based On Manifold Learning And Sparse Representation

Synthetic aperture radar(SAR) has been widely used both in civil and military application areas due to its great ability to work day and night under all weather conditions. SAR automatic target recognition(ATR) aims to realize target recognition based on target detection and orientation results from the radar echoes of the interested target and the background. SAR ATR extracts the features of the interested target, and then realizes its type, class, or configuration according to different demands under different conditions. It plays an important role in many crucial military application areas, such as command automation improvement, battle field interpretation, attack and defense abilities, homeland security, strategy warning, etc. It has been widely studied at home and abroad due to all the merits mentioned above.Based on the moving and stationary target acquisition and recognition(MSTAR) database, many advanced techniques like manifold learning, sparse representation and statistical modeling have been employed into the key step of the model-based SAR ATR algorithms in this dissertation, i.e. feature extraction. Algorithms have been proposed based on the advanced theories and techniques for SAR ATR, and satisfying results are obtained. The main works of this dissertation are summarized as follows:1. Considering the fact that the manifold learning algorithms are sensitive to the noise, a feature extraction algorithm that combines the locality preserving property and the Gaussian mixture distribution is proposed to realize SAR target recognition. The proposed algorithm extracts features in the view of statistics, and the useful information for recognition can be preserved as much as possible. The noise, or error of SAR images is described by a Gaussian mixture distribution, and the locality preserving property is embedded into the statistical model to preserve the local structure of the datasets. Experimental results on the MSTAR database validate the effectiveness of the proposed algorithm.2. Duo to the inherent speckle noise existed in SAR images, a feature extraction algorithm that combines the locality preserving property and the Gamma distribution is proposed to realize SAR target recognition. SAR images are described by the product model, and the speckle noise is modeled by the Gamma distribution to describe the essential characteristics of SAR images. And the locality preserving property is embedded into the statistical model to preserve the local structure of the datasets. Besides, we modify the affinity matrix to capture both the local and the global structure of the datasets for target recognition. Experimental results on the MSTAR database validate the effectiveness of the proposed algorithm.3. Due to the characteristic of the SAR image’s sensitivity to the target aspect angles, a Dempster-Shafer evidence theory based algorithm is proposed to fuse the detail and global features of the targets for recognition. The advantage of multiple sparse representations over sparse representation for detail feature extraction is analyzed. Two mass functions are constructed. One of them is established based on the multiple sparse representations to capture the detail information of the target, and the other one based on the sample statistical property is constructed to capture the global property of the target. The combined mass function has the advantages of both the detail and global features of the target. Dempster-Shafer fusion is carried out to achieve comprehensive description of the target for recognition. Experiments on the MSTAR database validate the effectiveness and superiority of the proposed algorithm.4. Considering the speckle noise in SAR images, a novel sparse representation algorithm using Bayesian fusion is proposed for SAR target recognition. The Bayesian fusion is employed to combine a better description of the SAR image and the sparsity. Due to the speckle noise in SAR images, the likelihood function is constructed by the Gaussian mixture distribution to model the statistical property of the noise term, and a better description of the SAR image can be obtained. The prior function is established by the Laplace distribution to guarantee the sparsity. The two functions are combined through Bayesian fusion. And the proposed algorithm has the advantages of the two functions and produces a more accurate sparse representation of the SAR image. Experimental results on the MSTAR database verify that the proposed algorithm performs better than several recognition algorithms previously presented.5. Since the sparse representation based recognition algorithm is worked in the sparse reconstruction space, we propose a novel dimensionality reduction algorithm. If the samples that are close in the high-dimensional space, using the proposed algorithm to reduce the dimensionalities of the samples, the corresponding sparse reconstruction of the samples in the low-dimensional space can still be close. The proposed algorithm can preserve the local structure of the samples in the sparse reconstruction space. Experimental results on the MSTAR database validate the effectiveness of the proposed algorithm.