Fuzzy Linear Discriminant Based On Cross Validation & Outlier Sample Processing And Its Application In Face Recognition

This paper focuses on studying the improvement of the traditional face recognition method based on linear discriminant analysis method (LDA). The linear discriminant analysis method (LDA) is a supervised classification method based on global statistical feature extraction. Due to the differences of the illumination angles, face positions, face occlusions and facial expressions lying in the training samples, when we classify the training samples using LDA method, this method may lead to inaccurate classification or incomplete extracting. To solve these problems, the training samples are divided into regular samples and outlier samples, using an improved fuzzy membership-based LDA method to classify regular samples and using Fuzzy C-means clustering method to classify outlier samples. Then the classification results of the two methods are combined to form the final classification result, this approach mainly takes into account different weight information, which is allocated by the contribution of extracting effective features, so overlapping information or outlier information of these samples has been taken into account in the feature extraction process, so that optimizing the results of the classification and improving the overall recognition rate of the algorithm.First, this paper studies the traditional LDA method from Eigenfaces to the calculation of the PCA, on this basis, C classification of LDA is deduced based on binary classification of Fisher LDA. Meanwhile, the concept of fuzzy membership and K-nearest neighbor algorithm combines to deduce fuzzy KNN algorithm used to calculate within-class scatter matrix and between-class scatter matrix in order to achieve optimal fuzzy projection matrix. What's more! It can optimizes the final classification result.Second, this paper introduces the improved fuzzy linear discriminant method. By proposing a relaxation normalization condition, we can use FKNN algorithm to calculate fuzzy membership (?ij)of the training samples. In the process of feature extraction, we use the contribution of the membership in scatter matrix redefinition. In this process, we need to predict the value of a number of parameters. Then we describe the method of cross-validation by which to predict the value of the parameters. One of the important functions of fuzzy membership (?ij) is to determine which sample belongs to regular samples or outlier samples. A discrimination method based on density estimation is described. By introducing Gaussian distribution function, we use 3a principle to judge which sample belongs to regular samples and which sample belongs outlier samples. After that, regular samples are processed in order to obtain a low-dimensional feature space which are used to get classification results.Third, this paper describes the outlier sample processing method based on C-means clustering algorithm, starting with the C-means clustering algorithm, this paper describes the method of fuzzy C-means clustering algorithm and the procedures using fuzzy C-means clustering algorithm to process outlier samples. The classification results of regular samples as a priori categories of information are used when we process outlier samples using fuzzy C-means clustering algorithm. This method improves the classification results of fuzzy C-means clustering algorithm.Experimental results based on AR and FERET database show that the proposed methods further improved methods to further improve the recognition rate.