Research On Online Dictionary Training And Weighted Discriminative Sparse Representation

Image processing is the major challenge of computer vision. The research of natural image understanding has greatly improved sparse coding algorithm. In recent years, image processing has achieved huge success with the development of dictionary training and sparse representation.At the aspect of dictionary training, sparse representation is a period of combination of all bases of a over-complete dictionary, utilizing non-unique coefficient vector as image representation and search for the most concrete coefficient vector as sparse representation of an image. There are two approaches to generate over-complete dictionary, one is randomly sampling training samples as dictionary, which widely applied in face recognition, another is to learn a over-complete dictionary through sparse learning model, which usually used in natural image classification. However, offline learning has to restrict the number of training samples, which decreases its classification accuracy. This paper proposed an online dictionary training algorithm, utilizing reconstruction error to update the current dictionary, greatly increases learning efficiency and realizes the real-time processing as well. But online learning has its instinct disadvantage of the decrease of classification performance. So this paper propose a novel Local Gabor Magnitude Weighted Phase (LGMWP) descriptor, and combine the SIFT descriptor to train a double dictionaries based on both descriptors. Experiments show that sparse representation based on double dictionaries outperforms offline learning.Recently, as the development of sparse dictionary learning, people begin to concentrate on the research of class discrimination and joint learning model of dictionary and classifier. Based on our analysis on state-of-art algorithms, it is figured out that those ones haven't fully used the reconstruction error in each class, which has huge discrimination and label information and will greatly affect the classification accuracy. As a result, this paper propose two approaches of weighted discriminative sparse representation, utilizing discrimination and structure in each class. Experiments prove that weighted discriminative sparse representation performs better than state-of-art approaches.