ROTATIONAL INVARIANT LOCAL FEATURES FOR TEXTURE CLASSIFICATION, A COMPARATIVE STUDY

Abstract

This thesis presents a comparative study of different rotation invariant local features for texture classification where classification accuracy of local features is examined and then compared with each other. Experiments are conducted on Outex datasets using Nearest Neighbor classifier. Thesis includes comparison of Local Features like absolute local difference, gray level of center pixels, standard deviation, mean, local binary pattern, and different combination of these features. All the methods are compared in terms of accuracy. Results of experiment have shown that although individually some local features may give poor result in classification but they can give enhanced results when used in combination with other local features. This study has helped us to conclude that gray level of center pixel when used in combination with absolute local difference and local binary pattern enhances the classification rate by adding information about center pixels which is not present in both of them individually. Combined feature of gray level, with local binary pattern and absolute local difference are compared with other techniques as well e.g. with invariant feature of local textures (IFLT) and gabor wavelets methods. It has been observed that our combined features give better results as compared to these techniques.