Nonlinear Scale Space Theory in Texture Classification Using Multiple Classifier Systems

Abstract

Textures have an intrinsic multiresolution property due to their varying texel size. This suggests using multiresolution techniques in texture analysis. Recently linear scale space techniques along with multiple classifier systems have been proposed as an effective approach in texture classification especially at small sample sizes. However, linear scale space blurs and dislocates conceptually meaningful structures irrespective of the type of structures exist. To address these problems, we utilize nonlinear scale space by which important geometrical structures are preserved throughout the scale space construction. This adds to the discrimination power of the classification system at higher scales. We evaluate the effectiveness of this approach for texture classification in Brodatz dataset using multiple classifier systems and learning curves. Compared with the linear scale space, we obtain higher accuracy in texture classification utilizing the nonlinear scale space.