Heterogeneous face image matching using multi-scale features

Heterogeneous Face Recognition (HFR) refers recognition of face images captured in different modalities, e.g. Visual (VIS), near infrared (NIR) and thermal infrared (TIR). Although heterogeneous face images of a given person differ by pixel values, the identity of the face should be classified as the same. This paper focuses on NIR-VIS HFR. Light Source Invariant Features (LSIFs) are derived to extract the invariant parts between two types of face images. The derived LSIFs rely only on the variation patterns of the skin parameters so that the effects generated from light source can be largely reduced. A common feature extraction method is designed to capture LSIFs based on a group of differential-based band-pass image filters, and we show that the scale for filters is critical. Our results in CASIA HFB database validate the effectiveness of the model and our recognition approach.