Supervised Dictionary Learning Based on Relationship Between Edges and Levels

Abstract

Categories of images are often arranged in a hierarchical structure based on their semantic meanings. Many existing approaches demonstrate the hierarchical category structure could bolster the learning process for classification, but most of them are designed based on a flat category structure, hence may not be appreciated for dealing with complex category structure and large numbers of categories. In this paper, given the hierarchical category structure, we propose to jointly learn a shared discriminative dictionary and corresponding level classifiers for visual categorization by making use of the relationship between the edges and the relationship between each layer. Specially, we use the graph-guided-fused-lasso penalty to embed the relationship between edges to the dictionary learning process. Besides, our approach not only learns the classifier towards the basic-class level, but also learns the classifier corresponding to the super-class level to embed the relationship between levels to the learning process. Experimental results on Caltech256 dataset and its subset show that the proposed approach yields promising performance improvements over some state-of-the-art methods.