Comparison Of Feature Extraction And Classification Methods For Face Recognition

Along with the rapid development of information science and biotechnology, face recognition is used in more and more fields. This thesis mainly studies and compares the performances of feature extraction and face recognition methods for the classical Olivetti Research Laboratory and Yale databases.For feature extraction, principal component analysis (PCA), linear discriminant analysis (LDA) and Fisherface methods are studied. We add the gray statistical characteristics and then make PCA. Compared with the direct PCA, this proposed method has relatively low time complexity and high recognition rate. This thesis proposes a kind of (2DLDA)2on the basis of2DLDA. The experimental results shows that (2DLDA)2can not only reduce the feature dimensions effectively but also have higher recognition rate and lower time complexity than the common LDA does.For classification models, the nearest neighbor (1-NN) classifier, neural networks and support vector machines (SVMs) are studied. We use the dimensionality-reduced datasets to train and test the performances of the three classifiers. We find out the following facts:(A)1-NN classifiers have the highest recognition rate among the three classifiers. But in order to label each test sample, a1-NN classifier has to compute the distance between the sample and all the training samples, and thus has the largest time complexity.(B) SVMs often have high recognition rate and low time complexity when establishing the separating hyperplanes by using the one-against-all decomposition mode. Under the circumstances, SVMs are a bit lower in the aspects of recognition rates than1-NN classifiers, but has the lowest time complexity among the three classifiers.(C) Single-hidden-layer neural networks may have relatively fast learning speeds because the numbers of training samples in the face databases are small. As a result, the neural networks have quite low training but high test error rates.