Research On Face Recognition Technology

Face recognition is one of most challenging problems in the fields of pattern recognition and machine vision. It also becomes an active research topic recently. But to date, it is still far from satisfactorily solved.The main contributions of this paper are as follows.I.A new method of feature point detection based on the geometric characters of face organs is presented. We segment the face image into several "Windows"according to the prOportion relationship of organs and detect the feature points in each window. In this way, the complexity of the algorithm is reduced and the correct rate of detection increases. Considering the importance of the face contours in recognition, we present a method of contour detection based on Canny抯 edge detection algorithm and the mathematical morphology algorithm. We use 6 feature points to describe the face contour. Finally, we implement face recognition using the method of feature point matching.2.A new method based on Singular Value Decomposition(SVD) is presented and we name it as local Singular Value(SV) based method. Traditional SV based method uses the SV features of the whole image. The method has an inherent deficiency that comes from the fact that only a few dominant SVs are used as characters of the image. The number of dominant SVs of image matrices is usually only a few. Increasing the size of image does not necessarily leads to an increase of the number of dominant SVs. Too few number of characters implies a difficulty in improving the recognition rate. To alleviate the difficulty, we propose a method based on local SV features. We use several "Windows"to partition the face image, calculate the SVs of each window, and rearrange the SV features into a new single SV feature vector according to the lexicographical order. In this way, the information of the image can be used sufficiently, and the number of domonant characters increases. Consequently, the individual difference among faces can be represented more properly. Experiments show that our method is more effective than conventional ones.3.A new method based on the Hidden Markov Model(HMM) is presented that uses the SV feature vector as the observation vector. We deem that a face should be described as a whole including not only each organ's numerical characters but also their various appearances and their relations. The rationality of using the HMM is implicated in the following idea. The model contains a set of hidden states that can be imaged as a set of distinct regions of the face image of a person. Various facial expressions and postures of the person are recognized as realizations of the set ofstates of the HMM for this person. For a different person, we use a different HMM. In this way, the numerical characters of organs of a person are associated with each other in a state transfer model. The 11MM method has been reported in the literature and the gray levels were used to characterize images. In our work, the SVs are used to characterize images. Owing to the noise-tolerance and transposeinvariability of SVs of image matrices, our 11MM outperforms the ones using the intensity values or the DCT coefficients. Our experiments verified the consideration.4.A method based on Embedded Hidden Markov Model(EHMM) and SV feature is presented. EHMM increases the states in horizontal direction and is a simplified 2D HMM. The EHMM can describe the face more elaborately and more completely. Our experiment using the ORL face database has acquired a recognition rate of over 99%. To our knowledge, this is also the best result using this database up to now.