Recognition, Hidden Markov Model-based And Multi-class Mapping

As a model of generative probability, Hidden Markov model has been widely used in image segmentation, pattern recognition and signal prediction for it simple and irrelevant to each other. The following tasks have been accomplished on the basis of previous work.1. The paper introduces the sub-space methods of facial recognition, especially for one-dimension principal component analysis, two–dimension principal component analysis, Fisher discriminant analysis and two-dimension Fisher discriminant analysis, and then discusses the recognition performance of algorithms on the number of training samples and feature dimensions. At last, in order to apply conveniently, the corresponding training time and recognition time is given as tables.2. The operation mechanism of Hidden Markov model is described deeply by means of mathematics statistics instead of the traditional method following to solve three problems. Combining with maximum likelihood estimation of probability function, it normally gets the re-estimate method of model parameters. All above making Hidden Markov model is easier to form a theoretical system. What is more, discrete cosine transform is replaced by wavelet transform in JPEG2000 image compression standard. So wavelet transform is introduced to abstract feature for Hidden Markov models to classification. Last but not least, features vector are generated following to fuse the features between the two and three from wavelet features, principal component features and two-dimension discrete cousin transform features, and then classified by Hidden Markov models. The experiments state a robust algorithm performance clearly with feature fusing strategy.3. The dual embedded Viterbi decoding algorithm of pseudo two-dimensional Hidden Markov models is depicted in detail. Then the pseudo two-dimensional Hidden Markov models are applied to face recognition with the features extracted by wavelet transform instead of discrete cosine transform. Besides, the more robust face recognition system is realized by fusing the wavelet features and discrete cousin transform features and classifying with pseudo two-dimensional Hidden Markov models. At the same time, the training time and recognition time of models is recited in detail for engineering application.4. Combining Hidden Markov model and discriminative method such as support vector machine enables us to construct more discriminant Fisher score features following to Fisher Kernel. And then the the principle of multi-class mapping is exploited to concatenate Fisher score to generate features. Taking use of the above features for face recognition achieves a better accuracy rate comparing the traditional linear discriminant methods and Hidden markov models. What is more, the features discriminating power of different parameters of models are discussed as well as how to choose the numbers of multi-class mapping to make sure a better price quality.