An improved adaptive discriminant analysis for single sample face recognition

Face recognition is an automated process with the ability to identify individuals by their facial characteristics. Currently there is a problem in which the process requires several examples of the person of interest's face in order to produce accurate outcome, and the process is intolerant to the variation in facial expression and the condition of lighting of the face image needed to be identify. This inspired us to come up with an algorithm to increase accuracy of single sample facial recognition process. In the case where multiple samples are available, the best approach to identify a person by face recognition system is to use Fischer Linear Discriminant Analysis (FLDA) method which use multiple samples to calculate the within-class scatter matrix and could give output accurately. However with only one sample it means the sample does not have any variation, hence impossible to find the within-class scatter matrix. The Adaptive Discriminant Learning (ADL) [1] was proposed to solve the problem by deducing the within-class scatter matrix from auxiliary generic set which consist of multiple samples per person then use FLDA to recognize face image. In this paper, we improve the method by preprocessing the input image using a local illumination normalization to make the feature of the face became more obvious and suppress the effect of illumination variation and incorporating a part-based methodology to further increase the recognition rate. The system was tested with the FERET face database, and the recognition rate is improved from 77% to 93%.