International Journal of Computer Science Issues

More than a traditional journal...

In this paper, an alternative approach for texture classification using an invariant texture representation and a tree matching kernel is proposed. The approach identifies regions of a given texture image using a Speed-Up Robust Feature or SURF descriptor. The regions of all training texture images are then clustered into a tree of non-uniformly shaped regions based on the distribution of them using a hierarchical k-means algorithm. The tree structure forms a tree of keypoints to be used for determining similarities between two texture images. The similarity is computed based on an approximate matching kernel called a tree matching kernel. Finally, Support Vector Machines (SVMs) with the tree matching kernels are constructed to classify textures. The performances of the proposed method are evaluated through experiments performed on textures from the Brodatz and UIUCTex datasets. The experiment results demonstrate that the proposed approach is quite robust to scale, rotation, deformation and viewpoint changes and achieves higher classification rates than some other well known methods.