Two-dimensional Clustering-based Discriminant Analysis for SAR ATR

This paper gives a new image feature extraction technique coined two-dimensional clustering-based discriminant analysis (2DCDA), which is based on 2D image matrices for constructing the scatter matrices and assumes that the data obeys the multimodal distribution. The detailed procedure of 2DCDA is to first partition each class of the data into multiple clusters via fast 2D global k-means clustering algorithm, and then try to find some directions such that the projections of every pair of clusters from different classes are well separated while the within-cluster scatter is minimized. Therefore, it fully exploits the cluster information and alleviates the linearly unseparable problem to some extent. According to the projection fashion, we divide 2DCDA into two versions, the right 2DCDA (R-2DCDA) and the left 2DCDA (L-2DCDA). They compress the image row or column only, so they need more features. To solve this problem, two-directional 2DCDA ((2D)
²
CDA) is developed, which compresses the image row and column simultaneously. Experiments have been carried out for recognition of three types of ground vehicles in the Moving and Stationary Target Acquisition and Recognition (MSTAR) public database to evaluate and compare the performances of the proposed algorithms with other methods. Results demonstrate that 2DCDA and (2D)
²
CDA are effective. And, the highest recognition rate is up to 97.89%, which is the best ever reported in the literatures.