Learning Based Reconstruction of Grayscale Face Image from Far-Infrared Image

Abstract

It is important for security surveillance systems to operate continuously for 24 hours. During the night, use of far-infrared cameras is preferable in outdoor situations due to a number of reasons. However, the person in the image is often unrecognizable. This paper attempts to estimate the face from his/her far-infrared image. The estimation is done through two phases, a holistic estimation and a patch based one. In each of these phases, a learning based approach is employed, which learns the relationship between grayscale and far-infrared face images from pairs of the images of a large number of persons. Canonical Correlation Analysis (CCA) is performed to obtain the maximum correlation in the data. Locally Linear Embedding (LLE) is performed to estimate grayscale face image. Three types of experiments were done with this method and evaluated by PSNR. These experiments show a good result in estimating face image whose face images of different expressions were included in training data.