Heterogeneous face biometrics based on Guassian weights and invariant features synthesis

Face images captured in different spectral bands are said to be heterogeneous. Although the heterogeneous face images from a same individual are significantly different in appearance, we can still achieve multi-modal patterns matching by image processing and transforming. In this paper, we propose a novel recognition algorithm based on face synthesis from NIR (near infrared) to VIS (visual light). For this first we use the illumination-invariant feature to construct face mapping function, then apply the correlation coefficient Gaussian kernel to determine the weights of synthesis components, and produce a synthesized VIS image corresponding to the query NIR image, thereby our problem is transformed to conventional homogeneous (VIS) face matching. Experimental results show that the proposed method effectively improves the recognition results.