Design and Analysis of A New Illumination Invariant Human Face Recognition System

摘要

In this dissertation we propose the design and analysis of a new illumination invariant face recognition system. We show that the multiscale analysis of facial structure and features of face images leads to superior recognition rates for images under varying illumination. We assume that an image I ( x,y ) is a black box consisting of a combination of illumination and reflectance. A new approximation is proposed to enhance the illumination removal phase. As illumination resides in the low-frequency part of images, a high-performance multiresolution transformation is employed to accurately separate the frequency contents of input images. The procedure is followed by a fine-tuning process. After extracting a mask, feature vector is formed and the principal component analysis (PCA) is used for dimensionality reduction which is then proceeded by the extreme learning machine (ELM) as a classifier. We then analyze the effect of the frequency selectivity of subbands of the transformation on the performance of the proposed face recognition system. In fact, we first propose a method to tune the characteristics of a multiresolution transformation, and then analyze how these specifications may affect the recognition rate. In addition, we show that the proposed face recognition system can be further improved in terms of the computational time and accuracy. The motivation for this progress is related to the fact that although illumination mostly lies in the low-frequency part of images, these low-frequency components may have low- or high-resonance nature. Therefore, for the first time, we introduce the resonance based analysis of face images rather than the traditional frequency domain approaches. We found that energy selectivity of the subbands of the resonance based decomposition can lead to superior results with less computational complexity. The method is free of any prior information about the face shape. It is systematic and can be applied separately on each image. Several experiments are performed employing the well known databases such as the Yale B, Extended-Yale B, CMU-PIE, FERET, ATu26T, and LFW. Illustrative examples are given and the results confirm the effectiveness of the method compared to the current results in the literature.