Research And Application Of Maximum A Posteriori Classification Method Based On Kernel Method

Pattern recognition, one of the most important branches in Artificial Intelligence is gradually formed with the development of information and technique and can be applied in many fields. The nature of pattern recognition is to finish classification according to the specific algorithm. Classification can be divided into face recognition and text classification from the perspective of object. Face recognition boasting long research time and mature technique in pattern recognition, has gained attention of numerous researchers for many years. In the past days, the research on face recognition mostly focused on linear recognition algorithm and the updated algorithm based on the original algorithm, not getting expected result. In recent years, the rapid development of statistical theory and Kernel function has greatly accelerated the improvement of face recognition and face recognition has been put into the statistical theory with the further research. The object to be studied in the text classification is usually showed in explicit form so the study is easier compared with that for face reorganization. As the basic method in the pattern recognition, classification technique is widely applied so the research on the classification algorithm is of great significance.This paper is to explore another new multiply classification method which can be well adapted to the distribution of sample data on the basis of analyzing multiply classification method. The paper firstly elaborates the basic theory of multiply classification method, shows the importance of feature extraction to information extraction. Secondly, in order to solve the problem of linear classification failing to deal with nonlinear classification, Kernel function is introduced to inject the original data into the Kernel space. Relying on these classical algorithms, the classifier is obtained by distribution of probability density function. Lastly, obtaining the multiply classification methods based on probability density function. This paper focuses on the following aspects: Firstly, PKMAP is proposed after analyzing Parzen window in non-parameter estimation. The theory model which shows PKMAP's accuracy to probability density function with the problem of small sample size data is described in detail while PKMAP's shortcomings with high time complexity is mentioned. Secondly, the multivariate t distribution is expressed in parameter estimation, and then the algorithm TKMAP is proposed by combining the multivariate t distribution, Bayes theory and Kernel function together. The experimental results will be got in the text and facial data sets. Thirdly, TKMAP is applied to verify the performance of the facial images with heavy-tailed noises. Adding five heavy-tailed noise to facial images to test the performance of method.The paper makes a systematical and detailed explanation on the theory of classification algorithm and its principle, and then tests the algorithm through the international standard UCI data sets and four typical facial image data sets Yale and Umist, ORL and BioID. The experimental results from these data sets show the feasibility and the effectiveness of the proposed algorithm.