Research On Some Key Issues In Machine Vision Face Recognition

Face recognition is a very active research field. Face recognition system supported by machine vision technique and artificial intelligence is getting popular in diverse applications, e.g. access control, surveillance monitoring, security checking and retrieval in photo collections, intelligent interface in consumer devices and other automatic control systems etc. Much progress in the face recognition technology has been made over the past two decades. However, the accuracy of face recognition is notably suffered from the existence of variable lighting, pose, facial expression, aging and other random factors. Further upgrading the performance of face recognition system is still a challenge task. This dissertation deals with the following key issues in face recognition research field.(1) Normalization of variant illuminations and compensation of shadows.Illumination variation is one of the most important factors affecting accuracy of face recognition. It is quite often that the difference between photos come from same person under different illuminations is more significant than the difference between photos come from different persons under same illumination. To cope with the variation interference an effective illumination normalization processing approach is proposed. The magnitude of gray intensity in a pixel is equal to the surface reflection parameter multiplied by illumination intensity at the specified pixel position. The ratio of the gray intensities at two nearby pixels will be approximately equal to the ratio of surface reflection parameters at these two pixels since the illumination intensities change very slowly versus the geometrical position change if there is no shadow occurrence. At a specified pixel, the ratios of the gray intensity at this pixel to the gray intensities of its neighbor pixels within a surrounding local region can be calculated and collected. Then put all these ratios into a convolving operation calculation with a Gaussian like weight function, which is used for estimating the influence of different distances between the specified pixel and other neighbor pixels; and finally a synthetic ratio will be yielded. The synthetic ratio comprehensively represents the surface reflection property at this pixel position and is invariant under variable illumination. This synthetic ratio can be named eigen gray intensity. The photo formed by the eigen intensities can be used as illumination normalization photo. We make a further modification on the above normalization processing by adding an additional convolving processing with considering the influence of different distances in color space between the specified pixel and other neighbor pixels. An illumination shadow detection and approximate illumination compensation approach is also presented.(2) Construction of dedicated region gray difference features for human face detection.The common distribution pattern of gray intensities on human faces is an important basis for face detection processing. Region gray difference feature possesses robust ability in discriminating gray distribution.2001 Viola and Jones proposed a new face detection approach using Haar like region gray difference feature and got very success performance. However, the number of features used in Viola-Jones approach is very large. A new construction approach of dedicated region gray difference features for human face is presented in this dissertation. First, an analytical face dataset with diverse typical face appearance is built. A statistical analysis of the gray intensity occurrence in each pixel was implemented on the analytical face dataset. The ratio of the mean gray intensity to its standard deviation at a pixel represents the occurrence gray statistically. This ration can be used as a statistical gray index of the pixel at specified position existing in face photos. The regions with different gray ranges can be found by segmentation of the corresponding gray index image with different ratio (index) threshold. The dedicated region gray difference features can be built by selecting the regions from brighter regions and darker regions correspondingly. The dedicated region gray difference features possess very strong discriminant abilities in face detection since they were built by face oriented analysis statistically. In my research group practice, the face detection error rate of a prototype system using 28 dedicated region gray difference features plus one global gray distribution feature had been reached similar level as with Viola approach using thousands Haar like features.(3) Face recognition with fusion of Gabor feature face representation and Curvelet feature face representation.Information fusion can provide recognition system with more information cues and analytical aspects and is an effective approach in upgrading recognition performance. The fusions of Gabor feature representation and Coverlet feature representation both in score level and in feature level are investigated. A new score normalization method for score fusion and a new feature fusion approach named GCF (Gabor Curvelet Fusion) are proposed. In score fusion a new score normalization should be implemented first individually, in which different scores were normalized and referred to the same face cumulative occurrence rates. The fusion score is the weight sum of individual scores, the weighting relation is determined by SVM method. In feature fusion, the Gabor features and Curvelet features were implemented with a principal component analysis (PCA) and a linear discriminant analysis (LDA) first for dimension reduction and discriminant enhancement. Then, the correlation between enhanced Gabor features and Curvelet feature come from same person were analyzed by canonical correlation analysis (CCA) repeatedly and two new feature projection data series, which were close correlated, will be generated. The fusion features were formed by sum up these two feature projection series term by term correspondingly. The new formed features can be named GCF fusion features. The experiments result testing on MBGC dataset show that both score fusion and feature fusion can effectively reduce the recognition error rates. Feature fusion with GCF features provides better performance than score fusion based on weighting sum method. Comparison with individual Gabor feature mode or Curvelet feature mode, the fusion feature approach with GCF features can reduce error rates up to about 30%.In the last chapter of the dissertation, a brief summary of the above research work was presented. Face recognition is a rapidly developing technology; however, there are still many fundamental problems confronting the face recognition researchers today. Some important and promising directions are suggested for future research.