Illumination robust single sample face recognition based on ESRC

Abstract

The extended sparse representation classifier (ESRC) is one of the state-of-the-art solutions for single sample face recognition, but it performs unsatisfactorily under varying illumination. There are two main reasons: one is that the mutual influences of the reflectance and illumination in intra-class variant bases of the ESRC are extreme under varying illumination, the other is that the specific identity information of the face in the generic set constrains the performance of ESRC, since the specific identity information of the generic face is redundant and should be removed. In this paper, the additive strategy is introduced to ESRC, which can efficiently eliminate aforementioned two issues. Two additive models: one is the reflectance and illumination additive (R&L) model and the other is the high-and low-frequency additive (H&L) model, are introduced to ESRC to obtain two new methods: R&L_ESRC and H&L_ESRC. Each method can be solved by a combined L1-minimization problem, and the final classification is determined by the sum of two weighting residuals: the reflectance and illumination residuals (or the high-and low-frequency residuals). In our experiments, the wavelet denoising (WDM) model and the logarithmic total variation (LTV) model are employed to extract facial features in R&L_ESRC and H&L_ESRC respectively. The performances of the proposed methods are verified on the Extended Yale B, CMU PIE, AR, ORL, and self-built Driver face databases. The experimental results illustrate that the proposed techniques outperform previous approaches under both normal and variant illumination conditions.