| Face recognition is one of the challenging subjects in the areas of pattern recognition and computer vision with a wide range of potential applications, and it also becomes an active research topic recently. Although it is easy for most human observers to identify different human faces, automatically computerized face recognition is a very difficulty problem, which remains largely unsolved.For a human face, the significant facial regions (forehead, eyes, nose, mouth, chin) from top to bottom and the local characteristic regions (temple, left eye, right eye, etc.) from left to right can be described as the state sequences of E-HMM. Faces appearing under different conditions can be recognized as presentations of the sequences of states of the E-HMM for this person. One E-HMM encodes one person's face features. E-HMM can better describe the structure features of faces in two dimensions with global and local characteristics, so face recognition based on E-HMM has achieved better performance recently.This dissertation deals with the problem of constructing statistical models of face images using E-HMM. We have improved the classical parameter re-estimating algorithm for E-HMM. Especially, two E-HMM/ANN hybrid frameworks are proposed for face recognition, which extend and strengthen the face recognition performance of E-HMM. Four main contributions of the research work are listed below:(1) In view of the contribution entropies of different training photos to the final face E-HMM are different, a new weighted synthesis method for re-estimating E-HMM parameters is developed and described in detail. During re-estimating the E-HMM parameters, every training sample is represented by one E-HMM, the model parameters for every sample are obtained firstly, then the different model parameters were synthesized to one model through weighted method, and the weights are adaptively calculated in the training stage. After the training, one person's face is represented by one E-HMM parameters. The weight of one training photo can be viewed as the contribution factor of the photo to the end synthesis E-HMM parameters. So the weighted synthesis E-HMM can reflect the entropy of different photos and encode the face structural knowledge and local region feature more successfully than that of the E-HMM trained by conventional algorithm.(2) A modular based incremental training method is proposed so as to train the model whenever new sample sets are added to training sets. Firstly, the samples are divided into several sample set, every sample set is represented by one modular E-HMM, the modular E-HMM parameters are obtained through doubly embedded Viterbi training algorithm, and the necessary temporary parameters in the parameter re-estimating process are also saved for the use of next step. After every modular E-HMM is trained, the modular E-HMM parameters are combined into one E-HMM to represent one person. The advantage of the modularintraining method is that the E-HMM parameters re-estimating process has good ability of adaptation: when new sample sets are added to the training sample, the information of the new sample sets can be conveniently combined into the E-HMM, and the computational work is reduced. Besides, the modular training method provides an answer for the problem of choice initial E-HMM parameters.(3) A novel E-HMM/ANN hybrid network is proposed for face recognition, which combines embedded- hidden Markov model (E-HMM) and artificial neural network (ANN) within the hybrid architecture. Firstly, E-HMM is used to parameterize face image, the output likelihood of the E-HMM is encoded to form the input vector and is sent to the ANN. By taking advantage of the discriminative training of ANN, the weak discrimination of the Maximum Likelihood criterion can be improved, and the recognition performance can be improved by means of the learning ability of ANN.(4) A hybrid model based on E-HMM/ANN is proposed, which integrates the advantages of E-HMM with that of ANN, and the training algorithm o...
|
PDF Full Text Request