面向计算机视觉问题的图匹配算法研究与应用

Feature correspondence is a fundamental problem in computer vision, which lays the foundations for many important tasks, such as object recognition, 3D reconstruction. It can be well defined, and effectively solved by graph matching, which is attracting more research interests, driven by more powerful computation and storage abilities of modern computers and some other factors. Aiming at some key problems in graph matching algorithms and their applications, in this project we plan to carry out research on two levels, with the first one to extend our previous work, and the second one to explore new research directions. Specifically, the first one consists of three studies, which are respectively adaptive graph matching, adjacency tensor based matching between hyper-graphs with different sizes, and graph matching based lunar surface image processing. And the second one consists of the exploration on relations between gradient based optimization and spectral decomposition based optimization, and the exploration on time/space effective solutions for the matching between huge size graphs. In this project, we pay great attentions to both theoretical deductions and realistic applications, where the above lunar surface image processing and huge size graph matching problems both belong to application driven research.

计算机视觉领域中目标识别、三维重建等很多任务往往建立在已确定特征点对应关系的基础上，但如何实现鲁棒的对应实际上仍是一个有挑战性的任务。特征对应可以通过图匹配良好定义并有效解决，受计算机运算、存储能力的提高和其他一些因素的刺激，该研究方向受到关注。在此背景下，针对图匹配算法与应用中的若干关键问题，本项目拟开展两个层次的研究，第一层次拓展申请人前期工作，开展以下三项研究：(1)自适应图匹配、(2)基于邻接张量的不同规模高阶图匹配、(3)基于图匹配的月面图像处理；第二层次探索新的方向，包括：（4）探索基于梯度优化与基于谱分解优化的图匹配算法间的内部关联、（5）探索大规模图匹配的时间/空间可行性方案。在保留前期工作注重理论推导的特点基础上，本项目同时注重应用指导，上述研究内容（3）、（5）均从实际应用需求出发，反推相关算法设计与改进。

本项目针对计算机视觉与机器人视觉任务特点，提出多个图匹配核心算法，其中典型算法包括（1）系统的提出权重共同子图匹配的概念、模型与优化方法，并推广到外点数目自适应的情况，（2）提出基于邻接张量的高阶图匹配方法，成功将目前高阶图匹配O(N6)的存储复杂度降低到O(N3)。从研究成果角度来看，通过本项目的支持，共发表论文33篇，包括第一作者与通讯作者论文16篇。其中负责人第一作者SCI国际期刊论文8篇，包括IEEE Trans. NNLS、IEEE Trans. Cybernetics、Pattern Recognition、IEEE Trans. SMC: Systems等本领域顶级与权威期刊。在算法应用方面，除了本项目研究内容涉及的在月面图像处理方面的系列工作，项目团队还利用水下图像与月面图像的相似性，将相关算法推广到水下机器人视觉任务，并基于此组队参加国家自然科学基金委主办的2017年首届与2018年第二届水下机器人目标抓取大赛，连续两年获得单项第一。