Hierarchical Classification of Images with Sparse Approximation

Using image hierarchies for visual categorization has shown to have a number of important benefits. For instance it enables a significant gain in efficiency (e.g., logarithmic with the number of categories [1, 2]). Moreover, arranging visual data in a hierarchical structure echoes the way how humans organize data and enables the construction of a more meaningful distance metric for image classification [3] (see figure 1). However, a critical question still remains controversial: would structuring data in a hierarchical sense also help classification accuracy? While our intuition suggests that the answer may be positive, up to date no method have shown conclusive results that can demonstrate the correctness of this claim for the most general case of large scale databases. In this paper we address this question and show that the hierarchical structure of a database can be indeed successfully used to enhance classification accuracy using a sparse approximation framework. We propose a new formulation for sparse approximation problem where the goal is to discover the sparsest path within the hierarchical data structure that best represents the query object. Extensive quantitative and qualitative experimental evaluation on a number of branches of the Imagenet database [4] as well as on the Caltech 256 [2] demonstrate our theoretical claims and show that our approach produces the best categorization results (in term of a number of hierarchical-based distance functions) over a number of competing large scale classification schemes that do not exploit the hierarchical structure of the database.