Biometrics is widely used in a flexible and convenient way, because of its uniqueness and stability, This is especially true for image-based biometrics, such as iris recognition, palmprint recognition, and face recognition. Feature representation of biometric images plays an important role in a real biometric system. Feature representation aims to design or learn a set of compact and discriminative filters which can extract robust and essential characters from the images. It involves the design of feature descriptors, feature selection and feature mapping. This thesis focuses on multimodal biometric systems to develop related machine learning algorithms on local feature extraction, feature selection and feature mapping learning. Main contributions of the thesis are as follows: 1.For feature descriptor, the influence of parameters of ordinal measures is studied and summarized for texture image recognition. Image content based texture filter is proposed except for spatial differential filter.Combining the spatial ordinal measures with texture ordinal measures can improve the robustness and discrimination of features. 2. A robust regularized linear programming feature selection model is proposed as a global optimal sparse learning model. The large margin criterion is adopted to design the loss function, and the useful prior information of each feature is introduced to update the weight. Finally a non-negative sparse constraint is enforced to ensure the sparsity of features.Extensive experiments on iris and palmprint recognition are conducted to validate the effectiveness and generalization of the proposed algorithm. 4. A back-forward sparse learning algorithm is proposed to select the top k features, and a unified framework of heuristic regularized sparse learning algorithm is summarized. Two extensions of robust FloatSparse and nonnegative FloatSparse learning model are formulated to implement efficient and robust top k feature selection. 5. A Boosting-like sparsity regularized model is proposed with sample selection. The weights of samples are updated at each iteration, thus different styles of features are selected at different iterations. Finally the two-stage learning process can obtain both sparse and complementary features. 6. A joint piecewise linear feature mapping and selection algorithm is proposed to map the selected low-dimensional features to high-dimensional features. Through combining single feature mapping and pairwise feature mapping, ...
修改评论