Ontology-driven hierarchical sparse coding for large-scale image classification

Abstract

An ontology-driven hierarchical sparse representation is developed in this paper, which aims to support hierarchical learning for large scale image classification. Firstly, a two-layer ontology (semantic ontology and visual ontology) is built to organize large number of image classes hierarchically, where WordNet is used to construct semantic ontology and deep features extracted by Inception V3 are used to construct visual ontology (visual tree). Secondly, a novel algorithm based on Split Bregman Iteration is developed to learn hierarchical sparse representation, i.e., learning a shared dictionary and a set of class-specified dictionaries depending on the two-layer ontology. For multi-class image classification, a tree classifier is trained according to the two-layer ontology by using the hierarchical sparse representation. Thirdly, for a given test image, multiple paths are simultaneously evaluated to achieve optimal prediction of its class label. Our proposed approach has been evaluated over three benchmark datasets: ILSVRC2010, SUN397, and Caltech256, and the experimental results have demonstrated that our approach is better than the original joint dictionary learning method and can achieve better accuracy compared with other approaches which use handcrafted features.