Study On The Methods Of Face Recognition In Human-Computer Interaction

It is a key problem that how to improve the performance of the face recognition, which can be used as ID verification in human-computer interaction. This thesis mainly studies the methods of face recognition. The main contributions are as follows,1. The subspace-based face feature extraction methods are studied, and eight different algorithms are introduced: PCA, KPCA, 2DPCA, DLDA, KDDA, 2DLDA, Fast ICA, and KICA. The theories, steps, advantages and disadvantages of each are analyzed in detail.2. A modified two-dimension linear discriminant analysis (2DLDA) algorithm is proposed. After lighting compensation, the input face images are processed by singular value decomposition (SVD) perturbation and wavelet transform firstly, and then the conventional 2DLDA is used to extract the features, whose dimension is further reduced by PCA algorithm. Finally, the support vector machines (SVM) classifier is used for classification and recognition. Experimental results on ORL database show that the proposed method has better recognition performance than other PCA, ICA, and LDA based algorithms.3. A real-time single face recognition system under the simple background condition in human-computer interaction is completed. First, YCbCr color model is used to find the facial region. Then, the facial image is preprocessed by three steps: (1) transform the color image into gray scale; (2) make the lighting compensation by logarithm transform; (3) resize the facial region with the same size. And then the features are extracted by 2DLDA, and the nearest neighbor (NN) classifier is used for classification and recognition. Finally, a real-time face recognition system for video size is implemented. Experimental results on our own laboratory database show that the face recognition accuracy is more than 90%. 320×240...