Discriminant Coupled Subspace Learning for Low-Resolution Face Recognition in Image Sets

This study presents a discriminant coupled subspace learning for recognizing the low-resolution query face sets against the high-resolution gallery ones. Considering the practical face recognition system, the appearance of facial images contains a variety of orientations, lighting conditions, occlusions and image resolutions. We address not only the low-resolution problem but also these dramatic appearance variations. Each subject is represented by a pair of low- and high-resolution image sets to provide more appearance variances. Principal component analysis (PCA) is applied to reduce the dimension of the high-resolution image such that the similarity between the low- and high-resolution images can be measured but the discriminant information is ignored. Hence, a transformation matrix is formulated to obtain a discriminant coupled subspace where the similarity between low- and high-resolution image sets of the same subject is maximized while those of between subjects are minimized. Note that the canonical difference is applied to measure the similarity between the low- and high-resolution image sets. Experiments are conducted on the Yale Face database B and Honda UCSD Video Database.