Abstract: To solve the small sample size (SSS) problem of Fisher linear discriminant analysis (FLDA) when it is applied to image recognition tasks, a novel image feature extraction technique is proposed which is called two-directional two-dimensional maximum clustering-based scatter difference ((2D)2MCSD) discriminant analysis. In this method, the possible clusters for each class are first found, and then the between-cluster and within-cluster scatter matrices are constructed from the 2D image matrices based on these clusters. Finally, projection vectors are sought by taking the difference of between-class scatter and within-class scatter as the discriminant criterion. Thus the method can not only deal with multimodal distribution problems but also avoid inverse matrix calculation and SSS problems, and increase the efficiency of feature extraction. Moreover, the (2D)2 MCSD compresses image rows and columns simultaneously, thus overcoming the problem of too many features of 2DMCSD and Alternate 2DMCSD. Experiments on a moving and stationary target acquisition and recognition (MSTAR) public database demonstrate that the (2D)2MCSD is more efficient than some existing subspace methods. Furthermore, compared with 2DMCSD and Alternate 2DMCSD, (2D)2 MCSD achieves higher recognition rates with much less memory requirements.

Key words: target recognition, synthetic aperture radar, clustering-based discriminant analysis, maximum scatter difference (MSD), Fisher linear discriminant analysis(FLDA)