Image segmentation and adaptive superpixel generation based on harmonic edge-weighted centroidal Voronoi tessellation

摘要

In this paper, we extend the basic edge-weighted centroidal Voronoi tessellation (EWCVT) for image segmentation to a new advanced model, namely harmonic edge-weighted centroidal Voronoi tessellation (HEWCVT). This extended model introduces a harmonic form of clustering energy by combining the image intensity with cluster boundary information. Improving upon the classic centroidal Voronoi tessellation (CVT) and EWCVT methods, the HEWCVT algorithm can not only overcome the sensitivity to the seed point initialisation and noise, but also improve the accuracy and stability of clustering results, as verified in several types of images. We then present an adaptive superpixel generation algorithm based on HEWCVT. First, an innovative initial seed sampling method based on quadtree decomposition is introduced, and the image is divided into small adaptive segments according to a density function. Then, the local HEWCVT algorithm is applied to generate adaptive superpixels. The presented algorithm is capable of generating adaptive superpixels while preserving local image features efficiently.