A comparative analysis of neural and statistical classifiers for dimensionality reduction-based face recognition systems.

摘要

Human face recognition has received a wide range of attention since 1990s. Recent approaches focus on a combination of dimensionality reduction-based feature extraction algorithms and various types of classifiers. This thesis provides an in depth comparative analysis of neural and statistical classifiers by combining them with existing dimensionality reduction-based algorithms. A set of unified face recognition systems were established for evaluating alternate combinations in terms of recognition performance, processing time, and conditions to achieve certain performance levels. A preprocessing system and four dimensionality reduction-based methods based on Principal Component Analysis (PCA), Two-dimensional PCA, Fisher's Linear Discriminant and Laplacianfaces were utilized and implemented. Classification was achieved by using various types of classifiers including Euclidean Distance, MLP neural network, K-nearest-neighborhood classifier and Fuzzy K-Nearest Neighbor classifier. The statistical model is relatively simple and requires less computation complexity and storage. Experimental results were shown after the algorithms were tested on two databases of known individuals, Yale and AR database. After comparing these algorithms in every aspect, the results of the simulations showed that considering recognition rates, generalization ability, classification performance, the power of noise immunity and processing time, the best results were obtained with the Laplacianfaces, using either Fuzzy K-NN.