Investigating the effects of scale in MRF texture classification
Scott Blunsden · 2005
This work sheds the light on an important problem that faces real-world texture classification. That of incorporating textural information present at several scales and the robustness of classifiers to viewing distance and zooming. A Markov random field framework is considered and the Varma-Zisserman classifier (Varma and Zisserman, 2003) (VZ classifier) is used as a starting point due to its high rates of classification on some difficult datasets (the CUReT dataset for example). A region selector (the scale-saliency algorithm by Kadir and Brady (Kadir, 2002)) is incorporated in the VZ classifier to select 'salient' or significant areas in an image and use them for texture classification. The performance of this method on several datasets is discussed and analysed (namely the CUReT and the Brodatz datasets). The VZ classifier is then updated to include multiscale information and use that for classification, which improves the performance of the VZ classifier for the CUReT dataset but shows lower classification rates for the Brodatz dataset. The reasons relating to the type of textures present in each dataset are discussed. Finally, a discussion of real-world data classification is given with summary and future directions.