Robust image matching via local graph structure consensus. (June 2022)

• Record Type: Journal Article
• Authors: Jiang, Xingyu; Xia, Yifan; Zhang, Xiao-Ping; Ma, Jiayi
• Is Part Of: Pattern recognition. Volume 126(2022)
• Publication Date: 2022-06
• Subjects: Image matching -- Feature matching -- Mismatch removal -- Outlier -- Image registration; Pattern perception -- Periodicals; Perception des structures -- Périodiques; Patroonherkenning; 006.4
• DOI: 10.1016/j.patcog.2022.108588
• Languages: English
• ISSNs: 0031-3203

Abstract

Highlights: We propose a novel objective for mismatch removal problem. Our method solves for the inlier set based on local graph structure consensus. It has a closed-form solution with linearithmic time and linear space complexity. It performs well in terms of robustness and effectiveness in feature matching. It has the potential for promoting high-level vision tasks, like image registration. Abstract: Image matching plays a vital role in many computer vision tasks, and this paper focuses on the mismatch removal problem of feature-based matching. We formulate the problem into a general yet effective optimization framework based on graph matching by combining integer quadratic programming with a compensation term for discouraging matches, termed as Local Graph Structure Consensus (LGSC). Considering the local area similarity of those potential true matches, we design a local graph structure for preserving geometric topology, which contains a local indicator vector and a local affinity vector for each correspondence. The local indicator vector is utilized for edge construction, while the local affinity vector represents the match correctness of the nodes and edges between two graphs. In particular, the ranking shift with scale and rotation invariance is exploited to represent the node affinity. Ultimately, we derive a closed-form solution with linearithmic time and linear space complexity. Moreover, a multi-scale and iterative graph construction strategy is proposed to promote the performance of our method in terms of robustness and effectiveness. Extensive experiments on various real image datasets demonstrate that our LGSC can achieve superior performance over current state-of-the-art approaches.