The Research On Face Recognition Based On Hidden Markov Model

The visual information contained in face images is one of the most important biological feature which can distinguish different people. Since the features of direct, uniqueness and convenience, it has taken more and more people's attention. Automatic face recognition technology is an important component of the biometric identification technology, which is one of the most challenging problems for computer vision and pattern recognition, becoming an active research topic in recent years. It mainly contains two aspects: face detection and face recognition.Researchers have developed many face recognition techniques recently. They can be broadly divided into the following categories: based on still image, based on video series, based on 3D, etc. Since the variable plastic of human face and the influence of many factors in imaging process, there is still a considerable distance from practical application. This paper lucubrates the face recognition methods based on Singular Value Decomposition and Hidden Markov Model.The work in this paper including:(1) A novel face recognition method based on Singular Value Decomposition (SVD) is proposed, which we name it as local Singular Value (SV) based method. The traditional method uses the SV vector of the whole image as identifiers, which has two obvious deficiencies. First, the SV vector of the whole image reflects the statistic characteristics of the entire image, which is not enough to the details. Second, because the SV vector as identifiers is decided by the first notable values, the increasing of image size may not be able to increase the number of identifiable characteristics. The less of the characteristics leads to the lower of the recognition rate. Based on the above facts, we propose a face recognition method based on local SV vector, which sampling face image with several windows and denotes facial feature by several SV vectors. In this way, the information of the image can be used more sufficiently, and the effective identifiable characteristics are enhanced. So we can reflect face characteristics more plummy. The experimental results show that this method is extremely effective, and the method based on the local SV is far more excellence to the traditional method based on the SV.(2) We propose a face recognition method based on SVD and Hidden Markov Model (HMM), which adopts SV vector as observed vector. When holding the face as a target for modeling in face recognition research, we have to pay attention to both the overall information of each structure and the local detail information of facial features. The overall information describes the macro feature of faces, and the detail information of faces is the key to distinguish different face. A face should be described as a whole, not only the numerical characteristics of each organ, but it also should include the different idea and relationship of each organ. It is reasonable for us to choose the HMM. We can recognize the rich representation of the same person's facial expressions and gestures as a series of realization of the same state. They correspond to the same HMM, and we have different HMM for different people. When describing and recognizing human face with HMM, we are not solely using the numerical characteristics of each organ, but associating these characteristics with a state transition model. Owing to the stability and the transposes invariability of SV vector, it is better for us to use it as an observation vector than use gray values or 2D-DCT coefficients directly. We gives continuous HMM estimation algorithm under the multiple observation sequence condition to improve the classic Baum-Welch algorithm, which enhances the identify capability of face recognition.Our experimental results also demonstrate this theory, the face recognition method based on continuous HMM outperforms classic HMM face recognition method both on the computational data and the recognition rate.(3) We propose a face recognition method based on Singular Value Decomposition (SVD) and Embedded HMM. EHMM is a simplified two-dimensional HMM, which can describe human face preferably. The model describes the details of the information of facial local characteristics by the embedded state of super-state. It associates the local information with state transfer, and describes the overall macro-face information with the diversion relation between super-states. When describing and recognizing human face with EHMM, we are not solely using the facial local characteristics, but associating these local characteristics with probability transition model and creating a complete description of the macro. The facial expression and recognition method based on EHMM is more reasonable. We emend the continuous EHMM estimation algorithm on the basis of one dimensional HMM, which optimizes EHMM face model and improves the recognition rate of face recognition method.Our experimental results also demonstrate this theory, the face recognition method based on continuous EHMM outperforms classic EHMM face recognition method.