New feature extraction approaches for face recognition

• Authors: Vo Dinh Minh Nhat;Sungyoung Lee
• Venue: AI'05 Proceedings of the 18th Australian Joint conference on Advances in Artificial Intelligence
• Year: 2005

All the traditional PCA-based and LDA-based methods are based on the analysis of vectors. So, it is difficult to evaluate the covariance matrices in such a high-dimensional vector space. Recently, two-dimensional PCA (2DPCA) and two-dimensional LDA (2DLDA) have been proposed in which image covariance matrices can be constructed directly using original image matrices. In contrast to the covariance matrices of traditional 1D approaches (PCA and LDA), the size of the image covariance matrices using 2D approaches (2DPCA and 2DLDA) are much smaller. As a result, it is easier to evaluate the covariance matrices accurately and computation cost is reduced. However, a drawback of 2D approaches is that it needs more coefficients than traditional approaches for image representation. Thus, 2D approach needs more memory to store its features and costs more time to calculate distance (similarity) in classification phase. In this paper, we develop a new image feature extraction methods called two-stage 2D subspace approaches to overcome the disadvantage of 2DPCA and 2DLDA. The initial idea of two-stage 2D subspace approaches which consist of two-stage 2DPCA and two-stage 2DLDA is to perform 2DPCA or 2DLDA twice: the first one is in horizontal direction and the second is in vertical direction. After the two sequential 2D transforms, the discriminant information is compacted into the up-left corner of the image. Experiment results show our methods achieve better performance in comparison with the other approaches with the lower computation cost.