基于概率推理的图像特征匹配统一理论模型及其应用研究

Image feature matching is a very challenging yet fundamental problem in computer vision and pattern recognition, and it is also a prerequisite in a wide range of vision applications. The major challenges lie in that the various and complex image transformations lead to difficulties in designing a reliable matching algorithm, the high complexity of model optimization limits its applicability in many real-world problems, and the combinatorial nature of the matching problem often results in missing true correspondences. The goal of this project is to overcome these challenges by pursuing a novel uniform theoretical model for image feature matching. More specifically, we plan to analyze the spatial distribution properties of the true and false correspondences, and propose a theoretical model based on probabilistic inference to achieve adaptive multiple scene matching. To optimize the proposed model efficiently, we plan to study a fast implementation based on kernel sparse representation, and it is significant for solving the large scale matching problem in real time. Furthermore, we plan to investigate a novel strategy for putative correspondence construction based on the neighborhood topology invariance constraint, which is able to largely promote the number of true correspondence. Moreover, the software/hardware platform for experimental validation will be realized, where three vision applications including visual homing, face hallucination, as well as image fusion will be adopted to verify the matching performance. Through this study, we will develop a uniform theoretical method for adaptive multiple scene matching in high efficiency and high precision. It will provide theoretical supports and solutions for intelligent vision applications whose performance is largely determined by the matching results, such as vision navigation, object recognition, intelligent surveillance and retrieval, etc.

图像特征匹配是计算机视觉与模式识别领域中颇具挑战性的基础研究问题，也是解决多种智能视觉实际应用的核心技术。其研究难点在于图像形变模式复杂造成匹配算法设计困难，模型求解复杂度高制约算法实用性，以及匹配组合优化特性导致真实匹配易丢失等问题。本项目旨在提出一种新颖的图像特征匹配统一理论模型来克服这些挑战。通过分析真实匹配与误匹配的空间分布特性，提出基于概率推理的匹配理论模型，实现自适应的多场景匹配；研究基于核稀疏表达的模型求解快速算法，为高精准条件下的大数据实时匹配提供保障；研究基于邻域拓扑不变性约束的原始匹配构造策略，有效抑制真实匹配丢失；开发算法实验及验证平台，结合机器人视觉归巢、人脸超分辨率和多源信息融合等视觉应用实例分析算法性能。研究成果将形成一套自适应的多场景高效高精准图像特征匹配统一理论方法，为复杂环境下的智能视觉应用提供理论支撑和解决方法，包括视觉导航、目标识别、智能监控与检索等。

图像匹配是搜索并建立起同一物点在两幅或多幅图像上对应特征点的过程，是实现图像分析到视觉感知的重要纽带，其精度和效率制约着人工智能视觉处理的应用和发展。项目从图像特征匹配统一理论模型的构建、高效快速模型优化算法的设计、高召回率高准确率原始匹配的构造等关键问题展开研究，构建了一套基于概率推理的图像匹配理论模型，形成了多场景自适应的高效高精准图像匹配方法；同时，进一步围绕图像匹配应用展开研究，在图像融合、图像超分辨率、视觉闭环检测等取得关键技术突破，成果如下：1）建立自适应的多场景图像特征点匹配统一理论模型,解决了传统方法需根据不同场景事先设定对应模型，难以应对图像形变大或形变未知的场景，无法保证特征点精准匹配等系列问题；2）设计基于傅里叶基底的高效快速模型优化算法，克服传统正则化理论计算复杂度较高的问题，使算法的时间复杂度近似为线性，实现算法模型处理时间上实时性，体现高时效性；3）构造高召回率高准确率的原始匹配，这意味着在保证模型普适与求解高效的基础上，进一步抑制真实匹配丢失并提高原始匹配的准确率，实现了高精准匹配；4）在形成以上图像匹配高效、高精准理论体系的基础上，通过理论突破，研发出基于高效高精准匹配的多源图像融合、图像超分辨率、闭环检测新方法，提出配准和融合统一框架将源图像的高精度匹配和融合两个步骤合二为一，克服了传统光场超分辨率方法深度估计的缺陷，同时实现准确的闭环检测。研究成果在IJCV、IEEE TPAMI/TIP/TNNLS等中科院一、二区TOP期刊发表学术论文46篇，其中ESI高被引/热点论文18篇；出版专著1部；授权国家发明专利1项；获湖北省自然科学一等奖1项。成果突破了复杂环境下视觉感知的高效高精准图像匹配中的瓶颈问题，对于智能视觉领域中以匹配算法为核心的众多热点应用，给出了通用可行的方法模型。