Multiscale Image Analysis Based on Bidimensional Local Mean Decomposition

Multiscale image analysis provides important feature inputs for the further image processing. This paper proposes a new multiscale image analysis method called bidimensional local mean decomposition（BLMD） on the basis of local mean decomposition（LMD）. Firstly, BLMD uses 8-neighborhood operator to obtain local extreme points; In order to eliminate the influence of saddle points when searching the local adjacent extreme, this paper proposes an adaptive window-based search method to control the number of local adjacent extreme points; Finally BLMD calculates the smooth envelope estimation function and local mean function to generate the product function, which decomposes images into different scale components. The results on synthetic images and typical real-world images indicate that BLMD is effective for multi-scale image analysis. In comparison with bidimensional empirical mode decomposition（BEMD）, BLMD presents the more effective and fast image processing results. In addition, the parameter sensitivity analysis approves that BLMD shows robust performance in image processing. Finally, the reasonable ranges of some key parameters for BLMD are given in this paper.