Texture classification through directional empirical mode decomposition

Abstract

This work presents a method for texture classification through directional empirical mode decomposition (DEMD). Although there have been many filtering based techniques proposed for texture retrieval, problems of non-adaptivity and redundancy are still hard to solve simultaneously. As a technique being introduced into signal processing, empirical mode decomposition (EMD) is an adaptive and approximately orthogonal filtering process. To apply EMD to texture classification, we propose a new method of extending 1-D EMD to 2-D case called DEMD. The approach adaptively decomposes images into local narrow band ingredients-intrinsic mode functions (IMFs) and extracts their features including frequency and envelopes. To improve its classification ability the fractal dimensions of the IMFs are also considered. Decomposition of several directions is computed for rotation invariance. Experiments for textures in Brodatz set and USC database indicate the effectiveness of our technique.