Thermal-to-Visible Face Alignment on Edge Map

Abstract

This paper presents a novel thermal-to-visible face alignment method based on edge map. The alignment procedure is inspired by iterative closest points matching. However, iterative closest point (ICP) sometimes converges to the local minimum for edge face alignment. In this paper, pointwise distance, which refers to the linear combination of positional and edge local pattern, is proposed as it gives good initial estimation of the closest points matching in iteration. Another problem is that numerous spurious corresponding points will occur in the closest pairs. Therefore, the selection criteria for partial closest pairs are designed, according to magnitude and orientation of the displacement, to remove the spurious corresponding points caused by edge faces. The alignment performance is evaluated by the bias of facial fiducial points and heterogeneous face recognition. The experimental results, including intra-class and inter-class, are demonstrated to compare partial closest points with local pattern with ICP, manual labeling, and Gaussian fields criterion. The alignment effect of variable expressions and illumination is also tested. Besides, the wearing glasses problem is well solved by glasses replacement.