Modified Edge-Weighted Centroidal Voronoi Tessellation for Image Segmentation
Mengfei Li, Ying Xu, Hongwei Li · 2019
The edge-weighted centroidal Voronoi tessellation (EWCVT) algorithm has been proved to be very efficient for image segmentation, especially when segmenting images with strong noise. However, when the objects to be segmented consist of special or complex small-scale structures, e.g. the pores in a core CT image, the EWCVT algorithm might fail to identify them from noisy background. In this work, we propose two modifications for the EWCVT algorithm to extend its ability for segmenting objects full of small or complex structures. The proposed model and algorithm encode two kinds of prior information: local spatial coherence of the image pixels and local spatial coherence of small structures. The effectiveness and efficiency of the proposed model and algorithm will be demonstrated by various numerical experiments.