Recently there has been an increasing interest in image segmentation due to the needs of locating objects with high segmentation accuracy as required by many computer vision and image processing tasks. While image segmentation remains a research challenge, ;;superpixel;; as the perceptual meaningful grouping of pixels has become a popular concept and a number of superpixel-based image segmentation algorithms have been proposed. The goal of this thesis is to examine the state-of-the-art superpixel algorithms and introduce new methods for achieving better image segmentation outcome. To improve the accuracy of superpixel-based segmentation, we propose a colour covariance matrix-based segmentation algorithm (CCM). This algorithm employs a novel colour covariance descriptor and a corresponding similarity measure method. Moreover, based on the CCM algorithm, we propose a multi-layer bipartite graph model (MBG-CCM) and a low-rank representation technique based algorithm (LRR-CCM). In MBG-CCM, different superpixel descriptors are fused by a multi-layer bipartite graph, and in LRR-CCM, the similarities of the covariance descriptors of the superpixel are measured by the subspace structure. Besides, we develop a new over-segmentation, called superpixel association, and propose a novel segmentation algorithm (SHST) which is able to generate hierarchical segmentation from superpixel associations. In addition to those unsupervised segmentation algorithms, we also explore the algorithms for supervised segmentation. We propose a model for semantic segmentation, named ;;generalized puzzle game;;, by which the segmentation information contained in the superpixels can be integrated into the supervised segmentation.
PDF
download
878987797
028-85220240
