An improved coupled spectral regression for heterogeneous face recognition

Coupled spectral regression (CSR) is an effective framework for heterogeneous face recognition (e.g., visual light (VIS) vs. near infrared (NIR)). CSR aims to learn different projections for different face modalities respectively to find a common subspace where the samples of different modalities from the same class are as close as possible. In original CSR, the projection for one modality is supposed to be represented by the data from the same modality. In this paper, we show that not only the samples of the same modality, but also all samples from different modalities are useful to learn the projection. Based on this assumption, we propose an improved coupled spectral regression (ICSR) approach which assumes the projections are linearly represented by all samples. Moreover, in order to improve the generalization capability, the locality information among samples is considered during the ICSR learning. Experiments on PIE, Multi-PIE and CASIA-HFB face database show that the proposed ICSR enhances the heterogeneous face recognition performance compared with the original CSR and validates the effectiveness of the proposed method.