Adaptive Hierarchical Multi-class SVM Classifier for Texture-based Image Classification Song Liu, Haoran Yi, Liang-Tien Chia, and Deepu Rajan Center for Multimedia and Network Technology School of Computer Engineering Nanyang Technological University, Singapore 639798 {pg03988006, pg03763623, asltchia, asdrajan}@ntu.edu.sg Abstract In this paper, we present a new classification scheme based on Support Vector Machines (SVM) and a new texture fea- ture, called texture correlogram, for high-level image clas- sification. Originally, SVM classifier is designed for solv- ing only binary classification problem. In order to deal with multiple classes, we present a new method to dynam- ically build up a hierarchical structure from the training dataset. The texture correlogram is designed to capture spa- tial distribution information. Experimental results demon- strate that the proposed classification scheme and texture feature are effective for high-level image classification task and the proposed classification scheme is more efficient than the other schemes while achieving almost the same classification accuracy. Another advantage of the proposed scheme is that the underlying hierarchical structure of the SVM classification tree manifests the interclass relation- ships among different classes. 1. Introduction With the rapid growth in the number of images, there is an increasing demand for effective and efficient image index- ing and retrieval mechanisms. For large image databases, successful image indexing will greatly improve the effi- ciency of content-based image retrieval (CBIR). One at- tempt to solve the image indexing problem is by using im- age classification to get high level concepts [8]. In such systems, an image is usually represented by various low- level features and high level concepts are learned from these features. Support Vector Machine (SVM) has recently at- tracted growing research interest due to its ability to learn with small samples and to classify high-dimensional data. However, SVM itself is capable of solving only binary classification problem [1]. In order to deal with multiple classes, several schemes have been proposed, e.g. (1) one- against-rest, (2) one-against-one and (3) decision directed acyclic graph SVM (DAGSVM) [3]. In the one-against-rest scheme, for a n−class classification problem, n classifiers SV M i ,i =1, ··· ,n, which takes the samples of class i as positive samples and the rest as negative samples, are con- structed during training. At the test stage, the test sample is evaluated against SV M i ,i =1, ··· ,n and the SV M i which gives highest decision value as the predicted class is chosen. In the one-against-one scheme, one classifier SV M i,j is constructed for every pair of classes (i, j ). There are N (N − 1)/2 classifiers in all. During the test stage, the samples are evaluated against all the pairwise classifiers SV M i,j . The final decision of the class for the test samples are determined by a voting scheme. In DAGSVM scheme, same as the one-against-one, N (N − 1)/2 pairwise classi- fiers are constructed. In the test stage, a list with all class candidates is created. During each test with the SVM, the class candidate which is given negative label by the SVM is removed from the list. Since the SVM classifier is a binary classifier, it is nat- ural to organize the SVM classifiers in a binary tree struc- ture. At each node, the classes are divided into two sepa- rate subsets. Therefore, we propose a new scheme, adap- tive hierarchical SVM classification scheme, for multiple classes. This scheme is a binary SVM tree, where each node of the tree represents a SVM classifier. Both the train- ing and testing phases of the classifier are carried out in a top-down manner. By utilizing this scheme in high-level image classification, it can improve the classification effi- ciency while achieving similar classification accuracy com- pared with other schemes. The structure of SVM classifica- tion tree also indicates interclass relationships among differ- ent classes, which provides useful information for measur- ing similarities of high-level concepts. We also propose a new texture feature for high-level image classification. The original texton [5] histogram approach has been extended to texture correlogram to provide texture description for a high-level concept. The remainder of the paper is organized as follows. Sec- tion 2 describes the proposed hierarchical SVM tree classi- fication scheme. The extraction of the texture correlogram is given in section 3. Section 4 presents the experimental re- 1