Face Recognition Based On The Subspace Methods

The process of face recognition includes three steps: face detection, locating characteristics of face, face characteristics extraction and recognition. Among them, face characteristics extraction and the characteristics recognition are the most fundamental and important one. This thesis studies the theories and methods of FR (face 'recognition) systematically, focusing on subspace pattern recognition.In the part of feature extraction, PCA,SpPCA,LDA,KPCA,KLDA algorithms have been used to extract features. In subpattern-based principal component analysis (SpPCA), an original whole large dimensional pattern denoted by a vector is partitioned into a set of equally-sized sub-patterns in a non-overlapping way and then local information on each sub-pattern is extracted PCA, it has been revealed that SpPCA is faster and more efficient than PCA through experiment. By using kernel-based analysis method, these algorithms first nonlinearly maps the original input space to a high-dimensional feature space, where the distribution of face patterns is hoped to be linearized. Then the PCA and LDA method is introduced to derive the optimal discriminant vectors in the feature space. The algorithm avoids the problem existed in linear methods.In the part of classifier design, the article pays more attention to the research of radial basis function neural classifier in the application of face recognition. Improved clustering approach combined unsupervised learning with supervised learning has been proposed to determine the parameters of hidden nodes, which solves the problems of conventional k-means clustering. Through adjusting weight, computing error rate and modifying the parameters of hidden nodes, optimal results will be achieved in the learning procedure. The algorithm avoids the slow convergence speed and local minima existed in conventional method.Through training and testing the images in the ORL face database, we can conclude that the improved radial basis function neural classifier has the advantage of strong learning ability and fast convergence speed. The recognition accuracy is advantage to nearest neighbor classifier. So, the adopted algorithms have strong practicality.