基于结构先验的低质量视觉信息恢复研究

Low-quality visual data captured in different under-controlled conditions, is a terrible obstacle to further research and products based on high-quality inputs. Low-quality visual data restoration has been a very challenging yet fundamental problem in computer vision, multimedia and machine learning. This proposal aims to conquer the bottle-neck issues including the lack in effectiveness and sufficiency of structural priors, the high complexity of optimization algorithms, and the degeneration in handling complicated cases. Concretely, this proposal attempts to exploit and formalize potential structures existing in visual data, find ways to tightly approximate and effectively optimize those structure models，apply the proposed models and algorithms to specific tasks like low-light image enhancement, bad weather interference suppression and pollution reduction, as well as analyze the effectiveness of proposed structures and the convergence and complexity of designed algorithms. We aim to construct a (relatively) complete set of effective structural priors, pursue a novel uniform theoretical framework for improving the quality of acquired visual data, and boost the performance in real-world applications. In addition, it is expected to generalize the theory and techniques developed in this proposal to other data types and tasks, and provide a new perspective for better analyzing and designing deep neural networks.

成因各异的低质量视觉数据是以视觉信息为输入所开展的后续研究以及衍生产品的一大障碍。有效恢复低质量视觉数据是计算机视觉、多媒体和机器学习领域中一个颇具挑战性的基础研究问题。本项目围绕低质量视觉信息恢复中存在的结构正则缺乏充分性和有效性、执行算法的高复杂度以及复杂场景下性能欠鲁棒等瓶颈问题展开研究。具体研究内容如下：可视数据潜在结构先验知识的探索和理论模型设计；模型的紧致逼近和高效优化求解；结合低照度图像恢复、去恶劣天气干扰和污染恢复等具体应用，分析所提出的结构知识的充分性和有效性以及算法的收敛性和复杂度。通过本项目的开展，计划为视觉信息建立一套相对完备的结构先验库；形成一个基于结构先验知识的可视数据质量提高的统一理论方法；推进其在军事、商业、娱乐等相关领域的应用。另外，期望能够将结构知识泛化到更为一般的数据类型和任务上以及为深度神经网络的行为特性分析和设计提供新的思路。

低质量视觉数据的质量增强不仅为用户提供更美观的输出，而且可为下游任务提供更为可靠的输入。围绕低质量视觉数据质量增强，本课题主要展开了以下研究：1）可视数据潜在结构先验知识的探索和理论模型设计；2）模型的紧致逼近和高效优化求解；3）投入到具体任务验证所提出模型和算法的有效性。本项目提出了一系列的技术方案突破问题瓶颈: a）建模图像内容结构互斥性和互相似性，利用图像组件/图像块之间的关联提升图像恢复的精度；b）解耦图像多降质因素并分而治之，针对图像增强问题，降低模型复杂度，简化深度网络设计；c）利用多源/多模态信号的丰富信息取长补短，输出信息丰富视觉美观的融合结果。部分成果已经发表于领域顶级会议和期刊，包括CCF A/一区论文14篇，CCF B/二区论文3篇。除此之外，申请国家发明专利7项，出版著作1部，并开发低质量视觉数据恢复增强原型软件1项。