基于超像素稀疏表示的图像超分辨率方法研究

With the rapid developments of ultra-high definition display technologies in recent years，the demand for high spatial resolution images grows faster． It is important to reconstruct the high resolution image from the corresponding low-resolution image using digital image processing techniques, without updating the imaging equipment. Several image superresolution methods based on sparse representation have been proposed. However, the complex structures in an image can not be accurately represented by the traditional pixel-level or block-level sparse representation, and the prior information in the low resolution input image has not been used for over-complete dictionaries learning. To solve those problems, this project focuses on investigating image superresolution method via sparse representation with superpixel. Combining sparse representation theory with superpixel segmentation method and multiscale analysis theory, firstly, we will propose a new superpixel representation method of complex structures in an image. Secondly, we will present a new superpixel-level dictionary learning algorithm with content and scale adaptive. Thirdly, we will design a structural sparse representation model with superpixel, and a prior model between the high-resolution superpixel samples and low-resolution superpixel samples. Finally, we will propose a new image superresolution method via sparse representation with superpixel to accurately reconstruct fine details, and improve the robustness of image superresolution method. The achievements can be applied to superresolution of the low-resolution images from painted murals.

超高清显示技术的快速发展必将对高空间分辨率的图像产生巨大需求，受成像工艺、成本和环境等因素制约，高空间分辨率的图像难以广泛获取，亟待研究能有效提升现有低质量图像空间分辨率的方法。虽然基于稀疏表示的图像超分辨率取得了一些研究成果，但是采用像素或图像块的稀疏表示方法对图像复杂结构表示不够准确，且过完备字典未能充分利用图像自身信息和样本图像先验信息。因此，基于稀疏表示的图像超分辨率方法仍有很大性能提升空间。 本项目将深入研究基于超像素稀疏表示的图像超分辨率方法，包括提出图像复杂结构的超像素表示方法，提出内容和尺度自适应的高-低分辨率超像素字典学习方法，建立高-低分辨率超像素间的关系模型，提出基于超像素的结构化稀疏表示模型和求解算法。突破像素或图像块等方式对图像复杂结构超分辨率的限制，精确重构低分辨率图像缺失的细节信息，提升图像超分辨率方法的性能，在低分辨率壁画图像的高清展示中取得实际应用。

本项目针对图像超分辨率过程中存在的严重病态问题，深入研究基于稀疏表示的图像超分辨率方法，融合局部和非局部约束、超像素表示方法和时空约束模型，提出多种图像复杂结构表示方法，改善图像复杂结构与内容的准确表示能力；设计多种内容与尺度自适应的过完备字典与学习方法，提升过完备字典学习方法对先验信息的获取能力和针对性；构建两类图像复杂结构高-低分辨率之间的回归关系模型，提出多种基于稀疏表示图像超分辨率模型和求解算法，精确重构低分辨率图像缺失的细节信息，提高基于稀疏表示图像超分辨率方法的准确性和鲁棒性。本项目取得了一系列有意义的研究成果，丰富了稀疏表示学习理论与方法，拓展了稀疏表示理论在图像超分辨率等领域中的应用研究，促进了基于稀疏表示的图像超分辨率技术发展。