The Research Of Face Recognition Based On Independent Component Analysis

Human face recognition is attractive in pattern recognition and image processing. It can be applied to security system, human ID management, teleconference, digital surveillance and so on.Generally, face recognition consists of three parts: preprocessing, feature extraction and classification. Feature extraction is the most important thing, on which our study focus.The preprocessing work includes image enforcement, geometry normalization of face images, image intensity normalization and whiting . These preprocessing improve image quality, decrease computation complex and therefore speed up algorithm' s convergence .In the feature extraction step, this paper uses independent component analysis to extract features. In the feature space, the components are statistical independent. The number of the independent component is determined by the number of input samples, so, when the train samples are very many, the training time will increase exponentially. Because of this, this paper decreases dimensions of the original input space by PCA, then uses PCA subspace as the input data of ICA algorithm, which reduces the number of input samples to ICA algorithm , decreasing the computation complex, so quickens the convergence of ICA algorithm .Considering the speciality of face images, We adopt the improved Infomax algorithm, which uses fourth cumulation to study gi(yi) adaptively.As ICA face feature space is concerned , this paper compares the SVM classifier with Nearest Neighbor classifier. The experimental results show ICA feature space is not sensitive to these two classifiers.At last, our experiments show that ICA is an effective method in face recognition field, taking on wide application foreground.