A robust diffusion approach to image segmentation based on curvelet enhancement

Abstract

In this paper, we present a new segmentation method in which Curvelet transform(CT) acts as an edge enhancement tool to modify diffusion marching. Firstly image segmentation is modeled via CT boundary emphasizing and lorentzian-function based diffusion. By means of multi-scale decomposition and multi-directional projection, CT detects pixels which are not obvious at pixel-level, but detectable by integrating over many pixels. Furthermore, projections inside Curvelet calculation directly lead to noise averaging, thus CT could be employed to retain weak edges and remove noises simultaneously when diffusion evolve to a certain extent. Secondly, a criterion is proposed to seek the appropriate moment for CT adoption during diffusion. It is fulfilled by analyzing histogram maxima every thirty iterations. If the count reduced between 2 maxima calculations arrives a threshold, CT will be performed to prevent edge disappearing. Thirdly, segmentation quality is measured to determine the cessation of diffusion. We carry out segmentation at the time when CT is completed or every fifty iterations finished. The partitioning numbers between two adjacent segmentations are compared to judge whether diffusion should be ceased. Experiments show that our approach takes CT's advantages of edge-preserving and denoising, it yields an efficient segmentation than the classical PDE does.