Date of Award


Document Type

Open Access Dissertation


Computer Science and Engineering

First Advisor

Song Wang

Second Advisor

Lili Ju


Propagated image segmentation is the problem of utilizing the existing segmentation of an image for obtaining a new segmentation of, either a neighboring image in a sequence, or the same image but in different scales. We name these two cases as the inter-image propagation and the intra-image propagation respectively. The inter-image propagation is particularly important to material science, where efficient and accurate segmentation of a sequence of 2D serial-sectioned images of 3D material samples is an essential step to understand the underlying micro-structure and related physical properties. For natural images with objects in different scales, the intra-image propagation, where segmentations are propagated from the finest scale to coarser scales, is able to better capture object boundaries than single-shot segmentations on a fixed image scale.

In this work, we first propose an inter-image propagation method named Edge- Weighted Centroid Voronoi Tessellation with Propagation of Consistency Constraint (CCEWCVT) to effectively segment material images. CCEWCVT segments an image sequence by repeatedly propagating a 2D segmentation from one slice to another, and in each step of this propagation, we apply the proposed consistency constraint in the pixel clustering process such that stable structures identified from the previous slice can be well-preserved. We further propose a non-rigid transformation based association method to find the correspondence of propagated stable structures in the next slice when the inter-image distance becomes large. We justify the effectiveness of the proposed CCEWCVT method on 3D material image sequences, and we compare its performance against several state-of-the-art 2D, 3D, propagated segmentation methods. Then for the intra-image propagation, we propose a superpixel construction method named Hierarchical Edge-Weighted Centroidal Voronoi Tessellation (HEWCVT) to accurately capture object boundaries in natural images. We model the problem as a multilevel clustering process: superpixels in one level are clustered to obtain larger size superpixels in the next level. The clustering energy involves both color similarities and the proposed boundary smoothness of superpixels. We further extend HEWCVT to obtain supervoxels on 3D images or videos. Both quantitative and qualitative evaluation results on several standard datasets show that the proposed HEWCVT method achieves superior or comparable performances to other state-of-the-art methods. vi