非结构化视点输入的光场重建技术研究

光场成像理论在显微观测、影视制作、辅助教学及裸眼三维显示领域具有重要应用前景，是计算摄像学领域最具潜力的发展方向之一。然而，现有光场成像技术存在采集系统复杂及成像质量低的问题，限制了光场成像的应用与发展。因此，本项目提出“非结构化视点输入的光场重建技术研究”这一课题，旨在从输入视点的非结构化空间分布几何特性出发，探寻具有强结构性约束的新型纹理模型对结构化输出光场进行表示，为实现高质量光场成像提供新的思路。具体研究内容包括：1）研究非结构化输入中，具有强结构性的新型深度纹理模型表示。2）研究新型模型表示中视点一致性特征的提取及表示方法。3）研究适合新型纹理模型表示的高质量光场重建算法。创新之处在于突破了现有光场重建方法各视点独立优化的计算框架，达到了重建高视点一致性的光场的目的。研究成果能够显著提升光场成像系统的成像质量和鲁棒性，部分前期成果已应用于2021年央视春晚直播中。

Light field theory has wide application prospect in microscopic observation, film production, auxiliary teaching and glass-free 3D display, and it is one of the most potential directions in computational photography. However, current light field imaging technology is suffered from the high complexity and low imaging quality of imaging systems, limiting the application and development of the light field theory. In this project, we propose the “Research on light field reconstruction from unstructured input views”, where a novel structured representation is presented to model the output light field by leveraging the geometry features of the poses in input and output views. Then, a high precision algorithm is proposed to reconstruct the light field from the structured representation. In this project, we firstly focus on the structured characters of the output light field and propose the structured representation for unstructured inputs. Then the view consistency of the novel representation is analyzed and extracted. Combining the novel representation and structured features, we propose a new method to reconstruct the light field. Different from traditional methods where each view is optimized individually, because the proposed method treats all views as a whole, the view consistency could be better preserved. This work could significantly improve the quality and robustness of light field imaging systems. Parts of previous work have been successfully applied in the CCTV New Year's Gala in 2021.