Fusion of bidirectional image matrices and 2D-LDA: an efficient approach for face recognition

Abstract

Although 2D-PCA and 2D-LDA algorithms obtain high recognition accuracy, drawback of these is that they need huge feature matrices for the task of face recognition. Besides, structure information between row and column direction cannot be uncovered simultaneously. To overcome these problems, this paper presents an efficient approach for face image feature extraction - a novel two-stage discrimination approach: preprocess original images to get two new image matrices and represent these images matrices using bidirectional 2D-LDA techniques. This approach directly extracts the optimal projective vectors from two new 2D image matrices by simultaneously considering row-direction 2D-LDA and column direction 2D-LDA. With this proposal, we can utilize the idea of local block features and global 2D images structures so it can preserve the 2D local facial features. Experimental results on ORL and Yale face database demonstrate that the proposed method obtains good recognition accuracy despite having less number of coefficient and few training samples (about two samples for each class).