An improved method for illumination invariant face recognition based on adaptive rescaling DCT coefficient in logarithm domain

This paper presents an improved method for robust face recognition using illumination normalization based on Discrete Cosine Transform (DCT) in logarithm domain. Two novel coefficients are designed to identify the lighting condition (LC), based on which the low-frequency DCT coefficients are adaptively rescaled except the first one (DC). As a result variations under different illumination conditions are minimized meanwhile original information contained in low-frequency is comparatively well preserved. Results of experiments on Yale B database and Extended Yale B database show that proposed method has better performance under variational input illumination conditions. The proposed method is fast in computation and could be easily implemented into real time face recognition systems.