IR face synthesis using motion vector field

Abstract

Infrared (IR) imaging device provides natural motivation for illumination-free face recognition. In this study, a method that can create a new IR face view between two given IR face images taken at different view points is presented, using motion vectors. Unlike existing prior arts which utilize optical flow, the proposed method utilizes the block based matching method to obtain correspondence between the two given IR face images. In addition, interpolation algorithm is employed to make the sparse motion vector "dense". An iterative error correction method is developed to reduce the errors of motion vectors. There is no facial features detection involved in our system so that it is suitable for IR images in which facial features like eyes, nose are blurred. The proposed method can be used for IR face recognition, animation, etc.