Research Of Texture Analysis Based On Empirical Mode Decomposition

Texture analysis is very important in remote sensing, medical image processing, computer vision and image searching by texture feature. Empirical mode decomposition (EMD), which is firstly introduced by Nunes et al. in 1998, can adaptively decompose a signal into several intrinsic mode functions (IMF) with the frequency lower and lower. This method, which analyses a signal completely by the inner scale feature directly from the data, is locally adaptive. It is widely applied in the signal denoising and fault diagnoses these years. Nunes et al. have extended this method into bidimension, and applied it in the texture analysis.The work of this paper mainly consists of three parts. Firstly, This paper analyses the appeared methods for avoiding boundary effect which is extended into bidimensional empirical mode decomposition, then an improved bidimensional empirical mode decomposition is presented in which the criterion for the 2D-sifting process to stop is modified, to avoid the early termination of sifting. Interpolation using compactly supported radial basis function is applied in the process of sifting, results in using different method for interpolation is analysed. Secondly, Bi-dimensional empirical mode decomposition is applied for classification of texture images. Texture classification uses the average and variance of envelops, average distances of points of local maxima and minima as features. At last, Texture images are decomposed using bidimensional empirical mode decomposition and segmented by the features which are obtained from the intrinsic mode functions. Features consist of the instantaneous amplitude and instantaneous frequencies from 4 directions.Modified fuzzy C-means algorithm are applied for the segmentation process. Some simulation experiments are carried out on software platform of MATLAB 7.0 and experiment results are analyzed and summarized at last. Texture images are decomposed into intrinsic mode functions by bidimensional empirical mode decomposition, then features for texture classification and texture segmentation are obtained and the experimental results are satisfying, which indicates that features proposed in this paper is reasonable.