Learning an Efficient Texture Model by Supervised Nonlinear Dimensionality Reduction Methods

This work investigates the problem of texture recognition under varying lighting and viewing conditions. One of the most successful approaches for handling this problem is to focus on textons, describing local properties of textures. Leung and Malik [1] introduced the framework of this approach which was followed by other researchers who tried to address its limitations such as high dimensionality of textons and feature histograms as well as poor classification of a single image under known conditions. In this paper, we overcome the above-mentioned drawbacks by use of recently introduced supervised nonlinear dimensionality reduction methods. These methods provide us with an embedding which describes data instances from the same classes more closely to each other while separating data from different classes as much as possible. Here, we take advantage of the superiority of modified methods such as "Colored Maximum Variance Unfolding" as one of the most efficient heuristics for supervised dimensionality reduction. The CUReT (Columbia-Utrecht Reflectance and Texture) database is used for evaluation of the proposed method. Experimental results indicate that the algorithm we have put forward intelligibly outperforms the existing methods. In addition, we show that intrinsic dimensionality of data is much less than the number of measurements available for each item. In this manner, we can practically analyze high dimensional data and get the benefits of data visualization.