Image classification: A novel texture signature approach

In this paper we present a novel image classification methodology based on texture signature. The approach consists of four distinct steps: 1) feature extraction from texture images without using any prior knowledge (e.g. viewpoint, illumination condition); 2) textures are modelled as texture signatures; 3) model selection and reduction is used to remove noise and outliers; 4) texture image classification using Columbia-Utrecht (CUReT) texture database. Classification performance was 91% accuracy for all 61 materials (2806 images) present in the CUReT database. The results are compared with texton based classification and effects due to various parameter settings are discussed.