Skip to Main Content
In face recognition applications, the dimension of the sample space is usually larger than the number of the samples in a training set. As a result, Fisher linear discriminant analysis (FLD) based methods suffers due to singularity problem (of scatter matrix). This situation is often referred as "small sample size" (SSS) problem. Moreover, FLD is a linear algorithm by nature. Hence, it fails to extract important information from nonlinear and complex data such as face image. To remedy this problem, this paper presents a new face recognition approach by integrating discrete cosine transform (DCT) and kernel Fisher discriminant analysis (KFDA). The DCT has the capability to compact the energy in an image and let the dimensionality of the input sample space to be reduced. Then, KFDA, a new variant of FLD, will be used to extract the most discriminating feature. This is performed by transforming the reduced DCT subset using a nonlinear kernel function to a high dimensional nonlinear feature space and then followed by the FLD step. Based on the extensive experiments performed on ORL database, the highest recognition accuracy of 95.375% is achieved with only 24 features.