基于时空结构自相似性与稀疏表示的地基空间目标图像复原

The ground-based images of space objects often encounter with the issues of heavy noise, saturation, and severe blur, which violate the traditional linear observed model and the assumed noise model. As a result, the traditional restoration methods perform weakly for real data thus cannot work well at all in practice. The project will develop nonlinear observed model and restoration method accounting for saturation pixels, wherein accurately extracting and modeling the prior knowledge are the key points. Here, two classes of characteristic knowledge, the multi-scale and multi-direction geometric structures and the inherent temporal signatures encoded by the observed image sequences are exploited. We propose to reveal the underlying characteristic knowledge by spatio-temporal structural self-similarity, and construct spare representation model for image restoration by means of statistical analysis in spatio-temporal-transformed domain and low-rank matrix representation. Basing on the models, we plan to handle the issues of heavy noise and saturated pixels in a combined way of dredging and blocking. To do this, some recovering methods combing spare representation model about spatio-temporal structural self-similarity are firstly studied to recover noise and saturated pixels, then the generation and propagation of residual saturation effects on deconvolution are analyzed, and subsequently some means are considered to prevent the spread of saturation effects. Furthermore, the point spread functions (PSF) are estimated in a progressive way with predicating and refining the edges from the blurred images. At last, the multiframe non-blind deconvolution methods in a sequential way with constraints of spatio-temporal structural self-similarity will be investigated so as to improve restoring the image details. In summary, the research will provide the theoretical methods and key technologies for the practical restoration of ground-based images of space objects.

地基空间目标图像存在强噪声、饱和与强模糊退化问题，经典的线性观测模型和噪声分布假设不再适用，传统方法复原真实数据能力差。本课题发展反映饱和因素影响的非线性观测退化模型与复原方法，先验知识的准确提取与数学模型表达是关键。挖掘空间目标序列图像蕴含的多尺度多方向几何结构与时域知识，提出采用时空结构自相似性揭示其本质属性，基于时-空-变换域联合统计分析、时空低秩矩阵，建立刻画时空结构自相似性的稀疏表示模型。在模型基础上，采用“疏堵结合”思路处理强噪声与饱和点：“疏”，利用帧内/帧间数据相关性，研究不依赖于噪声分布的强噪声与饱和点恢复方法；“堵”，分析反卷积过程中残余饱和影响的动态传播机制，研究阻止其扩散的处理手段。进而，研究强模糊边缘预测与多尺度渐进式PSF估计、时空自相似性约束的多帧递推非盲反卷积去模糊方法，提高目标细节恢复能力。研究成果将为地基空间目标图像复原实际应用提供理论方法和技术支撑。

本项目针对地基观测空间目标图像强噪声、饱和与强模糊等关键退化问题，围绕非线性观测与复原模型、特性知识建模及复原方法开展研究，取得理论方法成果：1）针对传统的线性退化模型不能准确建模强噪声、饱和点退化效应的问题，提出基于非线性退化模型的图像复原框架，自适应抑制反卷积迭代过程中强噪声、饱和点引起的振铃现象；2）针对多帧序列或多光谱图像时-空/谱-空结构联合建模难题，发现谱-空非局部自相似性最强、而非传统认为的时间维或谱间维局部相似性最强，提出低秩张量时-空/谱-空非局部自相似性结构表示模型，发展张量模向自适应低秩优化复原方法；3）针对空间目标图像暗背景平坦、亮目标结构丰富的特点，提出一种结合局部和非局部自相似性的空间目标图像去噪方法；4）针对传统方法难以准确建模未知混合噪声的难题，提出基于大数据深度学习拟合混合噪声分布的思路，发展一种两阶段卷积神经网络架构的混合噪声估计与复原方法；5）针对强噪声和强模糊耦合且先验知识建模困难的问题，提出一种模型优化-判别学习相结合的含强噪模糊图像反卷积复原方法，解耦强噪声、模糊退化并学习判别式先验信息；6）针对强模糊图像点扩散函数PSF难以准确估计的问题，提出基于边缘结构强度的PSF估计区域优选与模糊核尺寸预测方法，提升PSF估计的准确性。本项目执行期间，投稿学术论文16篇，其中已发表或接收13篇，包含计算机视觉领域顶级学术会议CVPR和ICCV论文2篇，遥感领域顶级学术期刊IEEE TGARS论文3篇、顶级学术会议IGARSS论文2篇，目前在Web of Science核心期刊中被引86次。申请发明专利5项，其中授权3项、受理2项。本项目完成了全部研究内容、达到了既定研究目标。