Infrared Face Temperature Normalization in Fourier Domain

This paper proposes a novel temperature normalization method in Fourier domain, which can lessen the effect on infrared face recognition from ambient temperature based on the idea of statistical learning in the transform domain. Firstly, the infrared face images in different ambient temperatures are transformed to Fourier domain. Secondly, based on statistical theory, the variances of phase spectrum and amplitude spectrum of the infrared face are used to describe the extent affected by the ambient temperature. Then, to achieve the robust information, those parts with big variances in the phase spectrum and amplitude spectrum are discarded and replaced by corresponding mean parts in training database. The main idea of this process is that one can set a suitable threshold for the variance of phase spectrum and amplitude spectrum and find those characteristic points that should be replaced. Finally, to verify the effectiveness of our temperature normalization method, the normalized infrared face can be applied to traditional face recognition system based on classic PCA method. Experimental results show our normalization method can get stable information in infrared face and improve the performance of the infrared face recognition system.