Study On Some Feature Extraction And Classification Methods For Face Image

Face Recognition has been one of the most active research topics in computer vision and pattern recognition. As an important biometrics technology, face recognition has and will have popular applications in the area of social public security and daily life. Now, there are still many key issues to be solved in face recognition, especially, the design of effective feature extraction methods and robust classifier. Therefore, this dissertation mainly focuses on face image feature extraction and robust classification methods. The major research results as follows:(1) A novel face image feature extraction method named discriminative histograms of local dominant orientation (D-HLDO) is proposed. Since the previous histogram of gradient based methods estimate the local gradients directly, the gradient orientations of these methods are not stable and sensitive to noise and brightness changes. To address this problem and further enhance the discriminative power of image features, this dissertation estimates the dominant orientation by using PCA based method and then provides the histogram of orientation in the local region. Finally, the local mean based nearest neighbor discriminant analysis (LM-NNDA) is used to get the low-dimensional and discriminative feature vector. The performance of the proposed method is better than other methods when dealing with noises and illumination changes.(2) A novel face image feature extraction method named local structure based image decomposition method (IDLS) is proposed. IDLS captures the structure information between central macro-pixel and its neighbors using the linear representation coefficients. One image is actually decomposed into a series of sub-images (also called structure images) according to local structure information. All the structure images, after being down-sampled for dimensionality reduction, are concatenated into one super-vector. Fisher linear discriminant analysis is then used to provide a low-dimensional and discriminative representation for each super-vector. Experimental results demonstrate that the proposed method is robust to the variations of illumination, expression and aging.(3)Based on the generalized Tikhonov regularization, a general regression and representation model (GRR) is proposed for face image classification. With the proposed model as a platform, this dissertation presents two concrete classification algorithms:basic general regression and representation classifier (B-GRR) and robust general regression and representation classifier (R-GRR). Most of the regularized coding based methods ignore the prior information hidden in the training set. To address this problem, this dissertation combines the generalized Tikhonov regularization, KNN and leave-one-out strategy to learn the prior information from the training set. Additionally, specific information is captured by using the iterative update algorithm. Experimental results demonstrate the performance advantages of GRR over state-of-the-art methods.(4) A low-rank regularized regression model is proposed for face image classification. Current regularized coding based methods convert an image into a vector at first. However, this operation ignores structure information of residual image. To address this problem, this dissertation introduced a rank function (which is replaced by a nuclear norm for easy optimization) of the representation residual image into a regression model. The model can be solved via the alternating direction method of multipliers (ADMM). Furthermore, we borrow the idea of robust regression and propose a robust low-rank regularization regression classifier. Experimental results demonstrate that the proposed method is robust to face recognition with occlusion, and yields better performances as compared to state-of-the-art methods.