Face Recognition Using Two-Dimensional Linear Discriminant Analysis (2DLDA)

Two-Dimensional Linear Discriminant Analysis (2DLDA) is a classical face recognition method that, compared to traditional one-dimensional LDA, directly processes two-dimensional image matrices. This approach avoids the high-dimensional problems associated with unfolding images into one-dimensional vectors. This method demonstrates excellent recognition performance on the ORL face database. The core concept of 2DLDA involves finding a projection matrix that maps original images to a low-dimensional space, ensuring that projections of within-class samples remain compact while between-class sample projections are well-separated. Since it operates directly on 2D image matrices, the computational complexity is reduced, and spatial structural information of images is better preserved. In implementation, this typically involves calculating between-class and within-class scatter matrices directly from image matrices rather than vectorized images. Experimental procedures on the ORL face database generally include several key steps: First, dataset preprocessing which may involve normalization and grayscale adjustment operations. Second, feature extraction using 2DLDA to compute optimal projection directions through eigenvalue decomposition of scatter matrices. Finally, classification using simple methods like nearest neighbor classifiers. Test results show that 2DLDA achieves high recognition rates on the ORL database, maintaining robust performance particularly with limited training samples. The algorithm implementation typically requires constructing image covariance matrices and solving generalized eigenvalue problems. Compared to other dimensionality reduction methods like PCA, 2DLDA places greater emphasis on inter-class discrimination capability, making it superior for face recognition tasks. However, it remains sensitive to variations in lighting conditions and poses. Subsequent research often combines 2DLDA with other feature extraction methods or deep learning techniques to further enhance performance, such as implementing hybrid approaches that integrate 2DLDA with local binary patterns or convolutional neural networks.