Research On Evaluation And Selection Of Feature In Pattern Recognition

Pattern recognition is an important part of information science and artificial intelligence. The major method of feature extraction and selection is to utilize certain mathematical tools to decrease pattern dimension in order to seek for the most effective feature-composition pattern vector with low dimension, which is applied to pattern recognition. The pattern recognition is rather crucial in the selection of feature. To some extent, the effect of feature selection refers to the performance and design style of grader. There are many standards to measure the feature extraction and selection. However, in the final analysis, most of them rely on the sorting algorithm and effect of the grader adopted. The paper gives a summy of feature selection algorithm. According to the formation of feature subset, all methods Feature selection algorithm can be divided into exhaustive methods, heuristic methods and random methods; according to evaluation function, it can be divided into filter and wrapper.The thesis starts with the perspective of statistical learning theory and applies the support vector machine algorithm to the assessment of feature selection. It provides the relevant definition (Correlation among the features, Dimensionality, Error sorting ratio, Positive class-among margin distance and Negative class-among margin distance) and assessment standard that affect the feature selection(Minimum error sorting ratio assessment principle, Maximum positive class-among margin distance assessment principle, Minimum negative class-among margin distance assessment principle, Minimum feature correlation assessment principle and Minimum dimension assessment principle). Finally, they are applied to the process of feature selection, and the effect of feature selection is improved. In practice, the assessment of the feature selection is a kind of feature selection method. Taking advantage of the assessment principles mentioned above can make further selection of the features, such as the feature correlation principle. In this paper, the evaluation of feature selection method is combined with genetic algorithm feature selection. And then we use an instance of the handwritten numeral recognition to illustrate the feasibility of the method.