Texture classification by combining local binary pattern features and a self-organizing map

Abstract

This paper deals with the combined use of Local Binary Pattern (LBP) features and a Self-Organizing Map (SOM) in texture classification. With this approach, the unsupervised learning and visualization capabilities of a SOM are utilized with highly efficient histogram-based texture features. In addition to the Euclidean distance normally used with a SOM, an information theoretic log-likelihood (cumlog) dissimilarity measure is also used for determining distances between feature histograms. The performance of the approach is empirically evaluated with two different data sets: (1) a texture-based visual inspection problem containing four very similar paper classes, and (2) classification of 24 different natural textures from the Outex database. © Springer-Verlag 2003.