Enhanced Diffusion by Projected Curvelet for Segmentation without Guiding Information

In this paper, we present a new segmentation model, which makes uses of Curvelet's advantages of edge preserving and noise averaging. The model first applies Lorentzian-function based diffusion for stable pixel clustering, and then projects boundaries by Curvelet transform (CT) to enhance edges and modify region smear in diffusion. In particular, we also propose a criterion to seek the appropriate moment for CT enhancement, it is fulfilled by comparing partition results of Lorentzian and Tukey-based functions. If the number of reduced regions between two adjacent segmentation rounds arrives a threshold, CT will be performed to prevent edge disappearing. Experiments show that this significant segmentation is resulted from CT's properties of boundary keeping and denoising, the method is superior to many other PDE approaches.