USING SUM MATCH KERNEL WITH BALANCED LABEL TREE FOR LARGE-SCALE IMAGE CLASSIFICATION

Abstract

Large-scale image classification is a fundamental problem in computer vision due to many real applications in various domains. A label tree-based classification is one of effective approaches for reducing the testing complexity with a large number of class labels. However, how to build a label tree structure with cost efficiency and high accuracy classification is a challenge. The popular building tree method is to apply a clustering algorithm to a similarity matrix which is obtained by training and evaluating one-versus-all classifiers on validation set. So, this method quickly become impracticable because the cost of training OvA classifiers is too high for large-scale classification problem. In this paper, we introduce a new method to obtain a similarity matrix without using one-versus-all classifiers. To measure the similarity among classes, we used the sum-match kernel that is able to be calculated simply basing on the explicit feature map. Furthermore, to gain computational efficiency in classification, we also propose an algorithm for learning balanced label tree by balancing a number of class labels in each node. The experimental results on standard benchmark datasets ImageNet-1K, SUN-397 and Caltech-256 show that the performance of the proposed method outperforms significantly other methods.