Quaternion matrix regression for color face recognition

TL;DR

This paper introduces a novel quaternion matrix regression method for color face recognition that effectively exploits color information and handles challenging conditions like occlusion and noise, outperforming existing approaches.

Contribution

It proposes a nuclear norm based quaternion matrix regression model and a robust extension using the logarithm of the nuclear norm for improved color face recognition.

Findings

Superior performance on public face databases

Effective handling of occlusion and illumination issues

Robust to mixed noise conditions

Abstract

Regression analysis-based approaches have been widely studied for face recognition (FR) in the past several years. More recently, to better deal with some difficult conditions such as occlusions and illumination, nuclear norm based matrix regression methods have been proposed to characterize the low-rank structure of the error image, which generalize the one-dimensional, pixel-based error model to the two-dimensional structure. These methods, however, are inherently devised for grayscale image based FR and without exploiting the color information which is proved beneficial for FR of color face images. Benefiting from quaternion representation, which is capable of encoding the cross-channel correlation of color images, we propose a novel color FR method by formulating the color FR problem as a nuclear norm based quaternion matrix regression (NQMR). We further develop a more robust model called R-NQMR by using the logarithm of the nuclear norm, instead of the original nuclear norm, which adaptively assigns weights on different singular values, and then extend it to deal with the mixed noise. The proposed models, then, are solved using the effective alternating direction multiplier method (ADMM). Experiments on several public face databases demonstrate the superior performance and efficacy of the proposed approaches for color FR, especially for some difficult conditions (occlusion, illumination and mixed noise) over some state-of-the-art regression analysis-based approaches.