基于自相似性和低秩表示的遥感图像去条带和去模糊联合复原方法

Stripe noise and blur are classical degradation problems in high-resolution remote sensing imaging, which greatly restrict the subsequent image processing and applications. Traditionally, image destriping and deblurring are researched separately. The project proposes to integrate these two problems into a unified variational framework, in which they can benefit greatly from each other. Modeling and representation of prior knowledge and the computational efficiency for image destriping and deblurring are two major bottlenecks to restrict their application. In this project, we analyze and study the structural characteristics of the stripe noise, remote sensing image nonlocal patch self-similarity sparse prior and low-rank matrix representation, and modeled these characteristics knowledge mathematically. Then multiple prior knowledge constraints are incorporated into a joint destriping and deblurring framework based on the theories of MAP estimation. Because of the introduction of the constraints for the gain change of the gradient across the stripes and self-similarity and low-rank constraints of the remote sensing images, the proposed restoration model can remove the stripe noise efficiently while restoring image details. To enhance the computational efficiency, a fast method called the alternating direction method of multipliers is introduced, and alternating iterative optimization and progressive implementation in a multiscale coarse-to-fine strategy speed up the solution of the unified recovery model. In summary, the research achievement will provide the theoretical methods and key technologies for the practical restoration of high-resolution remotely sensed images, and help promote the information processing in the field of high-resolution remotely sensed observation.

条带噪声和模糊是高分辨率遥感成像中常见的退化问题，严重制约后续图像处理与应用。传统上，去条带和去模糊问题被分开独立研究。本项目提出将去条带和去模糊整合在同一个变分框架下联合处理，并且两者可相互促进。遥感图像去条带和去模糊先验知识的建模与表达、计算效率问题是制约其应用的主要瓶颈。本项目分析和研究条带噪声的结构特性以及遥感图像非局部块自相似性稀疏性先验和低秩矩阵表示，并对这些特性知识进行模型化表达，建立整合多知识约束的广义MAP去条带和去模糊联合复原框架。提出的复原模型结合条带引起的梯度增益变化约束以及遥感图像自相似性与低秩约束，能有效去除条带噪声同时恢复图像细节信息。为了提高算法计算效率，引入交替方向乘子法，通过交替的迭代优化过程以及由粗到细的多尺度渐进式实现对联合复原模型进行加速求解。研究成果将为高分辨率遥感图像复原实际应用提供理论方法和关键技术，促进高分辨率遥感观测信息处理技术的发展。

针对图像中含有条带噪声、混合噪声和模糊退化问题，本项目开展了图像超分辨率增强处理的理论模型、方法算法和实验验证研究，主要研究成果如下:.(1)分析并研究了条带噪声的单方向特性，构建了非均匀性条带噪声校正算法。提出的方法引入迭代重加权最小二乘优化思路，同时构建了自适应正则化参数更新公式。提出的方法能有效地去除条带噪声，并且保存图像细节信息。.(2)提出了一种迭代重加权盲反卷积方法。提出的方法整合基于残差误差构造的自适应权进入数据项，使得提出的模型稳健于模型误差和异常值；整合图像在双边总变分算子作用下系数的空间自适应权进入图像正则化项，使得提出的模型在抑制混合噪声的同时更好地保存边缘和细节信息。.(3)提出了一种具有自适应正则化参数选取的盲反卷积模型。构建的盲反卷积模型的数据项中引入一种鲁棒回归机制，并且对图像施加总变分约束，对点扩展函数施加拉普拉斯正则化约束，最终构建的盲反卷积正则化模型能够有效地去除混合噪声同时保存图像细节信息。.(4)分析并研究了相邻谱带间谱信息的相关性和冗余性特征，我们构建了谱空总变分正则化模型，能去除高光谱图像的噪声和模糊，同时保持高光谱图像的空间和谱间不连续性，提高高光谱图像的质量。.(5)本项目提出的去条带方法可以应用到卫星遥感和红外成像领域，图像盲恢复方法可以应用到遥感和天文图像的分辨率增强。.项目组已完成项目研究内容，达到预期研究目标。本项目发表研究SCI收录期刊论文6篇。