An adaptive virtual sample generation method for one sample problem in face recognition
Jin Liu, A. Pengren, Hongbing Ji · 2016
Fisher linear discriminant analysis(FLDA) is a classical and important algorithm for face recognition. However, the FLDA will fail when there have only one sample each object, because the intra-class scatter matrices cannot be calculated. In this paper, an adaptive virtual sample generation method based on singular value decomposition(SVD) is proposed to solve one sample problem in face recognition. By using SVD, an approximation image is reconstructed, then, combining single training image and its approximation image to get virtual image. Every class has two samples: single sample and its virtual sample. The FLDA can be used to extract feature. The major contribution of the proposed method is that it adaptively construct virtual sample image based on energy distribution of different image. Experimental results show that the proposed method is efficient and have a higher recognition accuracy than based SVD and other existing algorithm.