Skip to Main Content
Most of the benchmark face recognition (FR) approaches are designed to depend upon the usage of holistic or texture-based features of the human face. Here we present a new approach to the problem of middle-wave infrared (MWIR) facial recognition that realizes the full potential of the MWIR band. It consists, first, of a fully automated standardization of MWIR images prior to feature extraction through: skin segmentation, eye detection, inter-ocular and geometric normalization of our entire face dataset. Then, a statistically-based physiological feature extraction algorithm is used that is tailored to MWIR phenomenology: infrared-based features are extracted that consist of wrinkles, veins, edges, and perimeters of facial characteristics using anisotropic diffusion and top hat segmentation. At the next step, fiducial points are detected either manually, or automatically using different detectors such as a fingerprint-based minutiae detector, the Scale-Invariant Feature Transform (SIFT) detector, and the Speeded Up Robust Feature (SURF) detector. Finally, face matching is performed, utilizing fiducial points originally detected: end points and branch points on the face are filtered using the maximum pixel distance allowed between two matching points. Matching experiments are performed by using either the whole or sub-regions of the human face. Facial matching results on holistic faces emphasize the importance of data pre-processing as we achieve a rank-1 accuracy of at least 95%, independent of the fiducial point extraction method employed.