Synthetic Aperture Radar Automatic Target Recognition Based On Projection Features

With the development of SAR (Synthetic Aperture Radar) technologies, SAR ATR (Automatic Target Recognition) has become an active research area. Because of the backscatters, SAR images are contaminated by speckle noise which will decrease image quality and hide the detail structure. Therefore, the effective feature extraction is one of the key steps for SAR ATR.SAR image feature extraction methods based on projection features are studied in this thesis. First, several typical methods such as Linear Discriminant Analysis (LDA),Primary Component Analysis (PCA) and Kernel PCA (KPCA) are discussed and analyzed. And then, Maximum Variance Projections (MVP) is studied in detail based on these methods. MVP is a linear discriminant algorithm that preserves local information by capturing the local geometry of the manifold. In the methods mentioned above, features are extracted in range subspace of samples. Finally, another feature extraction method based on orthogonal subspace of the samples is proposed. This method constructs projection subspace with the orthogonal subspace of samples'range subspace, and behaves good separability for SAR images which have similar structures. Moving and Stationary Target Acquisition and Recognition (MSTAR) dataset are used to test and prove the above two algorithms. Experimental results show that both methods are effective.