基于谱先验的图像空变盲去模糊非局部正则化方法

Space-variant blind image deblurring is a hot research topic in many applications, including video surveillance, remote sensing image, and medical imaging, etc. Solving this severely ill-posed task often requires regularization to yield high-quality results. First, this project establishes a robust model of noise based on the robust function. Then，we found that the blurry image itself encodes rich information about the blur kernel. Such information can be found through analyzing the spectrum property of the image as a linear operator which on the blur kernel. This analysis leads to a non-parametric priori model of the blur kernel which depends on the given blurry image. And then, there are nonlocal redundancies and diversity of locally structural features in the nature images. A nonlocal spectral priori model is proposed to fit different image contents adaptively by clustering the nonlocal similar patches and analyzing the cluster form of the induced spectrums. Next, use of K-L transform allows for modeling the space-variant PSF as a sum of orthogonal functions that are individually constant in form over the images, but whose relative amplitudes encode the PSF spatial variability. On the basis approach mentioned above, we propose a unified regularization and variational framework of both blur kernel estimation and image deconvolution. Finally，a fast and stable algorithm is designed to solve the proposed variational model. This project will enrich and promote the development of the modeling theory and algorithm in the field of image restoration, and the proposed theory and technique can be further extended to the many application fields such as image and video super-resolution reconstruction. Therefore, the study of this subject has important theoretical significance and practical value.

空变盲去模糊是视频监控、遥感图像、医学影像等应用领域的研究热点。该问题的高度病态性需要利用正则化方法对其进行求解。项目基于稳健函数建立鲁棒的噪声模型；挖掘观测图像中蕴含的模糊核信息，将退化图像视为作用在模糊核上的线性算子，研究该算子的谱特性，建立与观测图像相关的非参数化模糊核模型；挖掘图像中的局部结构特征多样性及非局部冗余性，通过非局部相似块编组，分析相似结构块的谱聚类形式，提出基于图像内容的自适应非局部谱先验模型；通过K-L变换将空变PSF分解为一组空不变正交分量的加权和, 加权系数由物空间PSF分布信息确定；基于所建立的噪声模型、模糊核模型及图像非局部谱先验模型，提出包括模糊核估计和图像反卷积的统一正则化变分框架；设计快速稳定的模型求解优化算法。本课题将丰富并推动图像复原建模理论和算法的发展，所提出的理论和方法可进一步推广到图像及视频的超分辨重建等应用领域，具有重要理论意义和实用价值。

图像盲去模糊属于图像处理和低层视觉中的关键性问题，是后续模式识别和高层理解的基础。近几十年来，该技术已经深入到遥感观测、医学影像、公共安全等多个应用领域。. 图像盲去模糊本质上属于数学中的不适定反问题，正则化方法是解决该类问题的有效途径。项目联合自适应统计学习和非局部正则化方法，提出了迭代优化重加权L1范数的图像盲去模糊算法；提出了基于提升的归一化稀疏性正则化图像盲去模糊方法；提出了基于亮度-颜色分治处理和边缘去振铃的CPU＋GPU并行加速盲去模糊算法；提出了彩色去马赛克的非局部鲁棒自适应稀疏表示方法；提出了彩色去马赛克的非局部张量联合稀疏编码方法；提出了联合空谱二阶特征的鲁棒彩色去马赛克方法；提出了图像泊松去噪的非局部TV正则化方法；提出了图像去噪的方向全变差正则化方法；提出了边缘和光谱信息保持的单幅图像超分辨率非局部重建方法。另外，将相关理论和方法推广应用于高维多通道图像的分类、多模态融合增强等，取得了前期性探索性研究成果。. 项目组合计发表论文17篇。其中国际期刊8篇，国内核心刊物4篇，国际会议5篇。SCI收录8篇，EI收录16篇(含SCI同时收录)。. 项目对于图像与视频的处理和智能分析具有广阔的应用前景，同时对于推动图像及视频的超分辨重建、高维信息理解和模式识别具有重要的理论意义和实用价值。