Down-Sampling Face Images and Low-Resolution Face Recognition

Though linear discriminant analysis (LDA) is popular in the field of feature extraction, they usually encounter two problems when applied to face images. The first problem is that the between-class and within-class scatter matrices of LDA cannot be evaluated accurately because their dimensions are usually much larger than the number of available image samples. The second problem is the small sample size (SSS) problem. However, if the face image can be resized into a small dimension, these difficulties may be overcome. With this paper, we analyze possible means to make LDA more feasible and effective for face recognition. Analysis and experiments show that down-sampling is very helpful for LDA to be performed with ease for face recognition. We compare a number of schemes that are used to exploit and combine information of the multi-level down-sampling results of the face images. We find that resizing conventional face images into smaller sizes may allow discriminant performance of LDA to be improved. There are two underlying reasons. The first one is that the face image of a lower dimension is very effective in helping LDA evaluate the between-class and within-class matrices more accurately. The second one is that LDA incline to obtain their best performance in an appropriate low resolution whereas the quantity of discriminant information what human beings can obtain is directly proportional to the resolution.