Discriminative Dictionary Learning And Its Application

Sparse representation is a powerful image representation model and has been widely used in many computer vision classification problems, such as image classification, face recognition and texture classification. Dictionary is an important factor which affects the final coding and performance of classification, so how to learn a discriminative and compact dictionary becomes a very challenging research work. Based on the summary and analysis of the research work of some related dictionary learning methods, this thesis proposes two kinds of discriminative dictionary learning methods and their applications from two aspects, including the geometry structure of the data space and the class information of the training data.This thesis proposes a locality-consistence-based discriminative dictionary learning method and applies it to image classification. It learns a discriminative dictionary and sparse codes using both geometrical structure of the data space which preserves the similarity between local features and the class information which makes the dictionary more discriminative. With the graph representation of samples, this method provides a promising paradigm for modeling the geometrical structure of the data space. In addition, the labels of measurements are imposed on the objective function. In this way, we can obtain a more discriminative dictionary such that close-by samples from the same class tend to have similar codes and ones from the different classes can take distinct codes with a large margin. This thesis applies it to some classical image classification datasets and experimental results demonstrate the effectiveness of our method.This thesis proposes a discriminative dictionary learning method based on Riemann manifold and applies it to face recognition and texture classification. In order to apply the data of symmetric positive definite matrix on Riemann manifold to sparse representation model, through the positive definite kernel function we map data and dictionary to the reproducing kernel Hilbert space in which each data point can be represented as a linear combination of based vectors in dictionary and measure the reconstruction error with Riemann metric. To consider the geometrical structure of Riemann manifold and class information, the codes of samples in the new space can keep high consistency. The similar matrix features from the same class in the original Riemann manifold can obtain similar codes and ones from different classes can get codes maintain difference. This thesis applies it to face recognition and textureclassification datasets. The experimental results on them demonstrate the effectiveness of our method.