Label group diffusion for image and image pair segmentation. (April 2021)

• Record Type: Journal Article
• Authors: Wang, Tao; Ji, Zexuan; Yang, Jian; Sun, Quansen; Shen, Xiaobo; Ren, Zhenwen; Ge, Qi
• Is Part Of: Pattern recognition. Volume 112(2021)
• Publication Date: 2021-04
• Subjects: Image segmentation -- Image pair co-segmentation -- Label group diffusion -- Geometry manifold -- Tensor product graph; Pattern perception -- Periodicals; Perception des structures -- Périodiques; Patroonherkenning; 006.4
• DOI: 10.1016/j.patcog.2020.107789
• Languages: English
• ISSNs: 0031-3203

Abstract

Highlights: Multivariate relationships are explored for segmentation task based on a unary probability conversion model. A multi-relational learning method is proposed to estimate the label group information on TPG, which imposes a higher-order graph Laplacian to smooth the prior label. A generalized LGD framework is established for semi-supervised image segmentation and unsupervised image pair co-segmentation tasks. Abstract: Diffusion technique is powerful for semi-supervised image segmentation since the geometry of the data manifold can be captured by the affinity propagation. Conventional diffusion methods focus on single label, which however ignore interactions among labels. This workstudies a generalized diffusion framework that considers label group for diffusion (LGD). The proposed framework can effectively capture the interactions among image elements via tensor product graph (TPG). A multivariate affinity framework is proposed to learn on higher-order TPG. Label pair diffusion algorithm is naturally derivedfrom the framework by considering second-order affinity (LG (2) D). We theoretically show that conventional label diffusion is the simplest case of the proposed framework (LG (1) D). Extensive experiments on image segmentation and image pair co-segmentation datasets demonstrate the superior performance of the proposed framework.