由于相机的景深有限，三维场景的照片常常因为局部散焦而出现模糊。散焦模糊的程度和景物与相机的距离紧密相关，因而通常是空域变化的，并且可以用于散焦图像复原、深度估计、图像重聚焦、前／背景分割、显著性检测、散焦放大等领域。本文研究基于单幅图像的散焦估计，即从一幅传统相机拍摄的可见光图像估计其各个像素点处因散焦而出现的模糊程度。根据散焦信息的分析方式，可以将散焦估计方法分为基于边缘的散焦估计与基于区域的散焦估计。本文针对当前这两类算法中存在的问题展开了相关研究，并给出了三个散焦估计的应用示例。 在调研了现有基于边缘的散焦估计算法后，我们发现它们所得的散焦图往往容易受到输入图像纹理的影响，并且其所用的边缘散焦模型都无法很好地描述结构边缘(两侧有深度跳变)的散焦过程。为此，本文提出了一种基于双参数散焦模型的散焦估计算法。在散焦边缘建模时，本文对边缘两侧分别使用一个参数来描述其散焦程度，不仅可以很好地描述结构边缘与纹理边缘(两侧深度相同)散焦过程的差异，而且为区分这两类边缘提供了可能性。基于此双参数散焦模型，本文设计了一套从再模糊的边缘中估计其两侧散焦参数的方法，并生成了纹理边缘置信度；然后此置信度被用于结构-纹理分解，在去除输入图像纹理的同时较好地保留了结构信息；最后以边缘点处的散焦估值为初值，以结构分量为指导图像，使用Laplacian抠图获得了对输入图像纹理不敏感的散焦图。 本文在简要介绍现有的三类基于区域的散焦估计算法后，针对基于频域分析方法中“散焦跳变与图像边缘不吻合”的问题，提出了一种基于改进似然特征和边缘线性基的单幅图像散焦估计算法。为了尽可能有效地提取散焦信息，本文分析了似然函数中各项的贡献，并提出使用两个互相垂直的梯度滤波器并分别用与之夹角小于45°的Gabor滤波器子集进行局部频域分析。在提取似然特征时，本文使用似然函数的前M个极大值点，在提取有效信息的同时降低了特征维度。在特征精化部分，本文使用回归树场并以基于边缘的方法作为线性基，成功地将基于边缘方法的优点整合进来。最终本文算法所得的散焦图既对输入图像的纹理不敏感，又较好地捕捉到了散焦跳变的边缘。 在散焦估计的诸多应用中，本文以散焦图像复原、前／背景分割和深度估计为例，说明了本文所提方法的用途。实验表明，本文所提的两个方法可以在这三个应用上取得令人满意的结果。 关键词：图像散焦估计；双参数散焦模型；局部频域分析；散焦图像复原

 Due to the limited depth of field of cameras, photos of a three dimensional scene are often distorted by image blur, which is caused by partially defocus. The amount of defocus blur is closely related to the distance from the object to the camera, therefore, it is usually spatially varying and can be used in many applications, such as defocused image restoration, depth estimation, image refocusing, foreground/background segmentation, saliency detection, defocus magnification, etc. This dissertation focuses on defocus map estimation from a single image, i.e., to estimate the blur amount caused by defocus for every pixel from a single visible image captured by a conventional camera. According to the way of analyzing defocus information, defocus map estimation methods can be classified as edge-based defocus map estimation and region-based defocus map estimation. This dissertation performs related researches aiming to fix the problems in the existing two categories of methods, and gives three examples for the applications of defocus map estimation. After reviewing existing edge-based defocus map estimation methods, we find that their defocus maps are usually influenced by textures of the input image. Besides, none of their defocus models can well describe the defocus of structural edges where there is a depth jump over the edge. To address these shortcomings, this dissertation proposes a defocus map estimation method based on a two-parameter defocus model. When establishing defocus model, this dissertation uses one parameter for each side of the edge to describe the defocus amounts. This can not only describe the different defocus situations for structural edges and textural ones (where the depth keeps the same over the edge), but also make it possible to distinguish these two kinds of edges. Based on this two-parameter defocus model, the dissertation designs an approach to estimate the defocus parameters for both sides through the re-blurred edges, and generates a textural edge confidence. Then the confidence is used in structure-texture decomposition, eliminating the textures while preserving the structural information reasonably well. Then the final defocus map, which is insensitive to textures of input image, is obtained through Laplacian matting, with the defocus estimations at edge points serving as initial estimations and the structure component serving as the guided image. This dissertation introduces the existing three categories of region-based defocus map estimation methods. After this, the dissertation proposes a defocus map estimation method using an improved likelihood feature and edge-based linear basis, to fix the problem in frequency analysis based methods that defocus discontinuities do not coincide with image edges. In order to extract defocus information as effectively as possible, this dissertation analyzes the contribution from each term of the likelihood function, and proposes to employ two orthogonal gradient filters in the localized frequency analysis, along with their corresponding Gabor filter subsets where the angle between the directions of the Gabor filter and the gradient filter is smaller than 45 degrees. When extracting the likelihood feature, the dissertation uses the first M highest local maximums of the likelihood function as the feature, extracting essential information while reducing the feature dimension. As for the feature refinement, this dissertation employs regression tree fields and uses edge-based methods as the linear basis, integrating the advantage of edge-based methods into our method successfully. The final defocus map is not only insensitive to textures of input image but can also catch the defocus discontinuities reasonably well. Among the numerous applications of defocus map estimation, this dissertation demonstrates the usage of the proposed methods on three applications, i.e., defocused image restoration, foreground/background segmentation and depth estimation. Experiments indicate that the two proposed methods can obtain satisfied results on these three applications. Key Words: defocus map estimation for image(s); two-parameter defocus model; localized frequency analysis; defocused image restoration