Distributed kernel Fisher discriminant analysis for radar image recognition
Jiansheng Fu, Wanlin Yang · 2011
In this paper, an extension of kernel Fisher discriminant (KFD), namely, distributed KFD (DKFD), is proposed for multi-class cases. Unlike the generalized discriminant analysis (GDA) which deals with all classes synchronously, DKFD uses a cost—normalized KFD for two-class problems and thereby covers the whole system by the KFD units. Based on DKFD, three new models, i.e., active DKFD (A-DKFD), passive DKFD (P-DKFD) and global DKFD (G-DKFD), are proposed for classification. Theoretically analysis and experimental results on radar HRRP databases indicate as follows. Firstly, compared with GDA, DKFD not only needs less computation time and space, but also is more convenient for multi-class cases and distributed computing. Secondly, in terms of recognition performance, the three proposed models, especially G-DKFD, surprisingly outperform GDA in general.