Research On Illumination Problem In Face Recognition

Face recognition, a popular modality of biometrics, due to its availability and relative uniqueness, is becoming more and more popular. Though study and practice show that face recognition is effective when used in controlled environment and strong progress have been made in the overall area of face recognition, there still exist obstacles for its widespread using. The variations of illumination is one of the issues obstruct using of face recognition, and has become a priority. Experimentally and theoretically, It has been shown that the difference between the images obtained from the same individual in different light conditions is larger than the difference between images obtained from different people in the same light condition.In recent years, empirical mode decomposition (EMD) which as a adaptive analysis tool of non-stationary signals has been employed in multiple applications such as image compression, image fusion, texture analysis. EMD decomposes the input signal into several intrinsic mode functions (IMF), which response to the intrinsic oscillating mode of the signal.while the main disadvantage of EMD is that it lacks a theoretical analysis. This paper use empirical mode decomposition to handle the issue of illumination variations in face recognition. The face images are decomposed into a collection of IMFs first by using EMD, then we can reduce the overall effect of illumination variation in face images by reconstructing the images by isolating these IMFs correspond to the illumination variations.In this paper, based on the essence of EMD, we select more proper methods for extreme point detection and interpolation in the sifting process. Firstly, the extrema of the subsignal having the higher instantaneous frequency,instead of the signal extrema, are used as interpolation points, Even if the extrema of the subsignal with the higher instantaneous frequency are not known in advance,this new interpolation points criterion can be effectively exploited in doubly-iterative sifting schemes leading to performance, and experiments shown that the performance of EMD was significantly enhanced by the using of the improved decomposition method.Then we expand the EMD to two-dimensional form also by selecting suitable methods for extreme point detecting in face images and appropriate interpolation algorithm and the face images are decomposed directly by the two dimensional EMD, while with the one dimensional EMD, the original face images must be converted into one dimensional form before they are decomposed. The experiments which we performed by using the algorithm of Principal Component Analysis (PCA) and EMD, show that EMD is more simple and effective as a method for the handling of illumination variations in face recognition compared with that of PCA.Finally, the whole work of this paper and the future research are concluded and discussed.