【作者】 张小红；

【导师】 童若锋； 唐敏；

【作者基本信息】 浙江大学 ， 计算机科学与技术， 2012， 博士

【摘要】 本文研究的超分辨率和图像去运动模糊是计算机图像处理领域一个非常具有挑战性的课题。理论上,超分辨率和图像去运动模糊都属于数学问题中的具有“病态性”的一类反问题。一个方程的“病态性”表现为该方程的解非连续地依赖于观测数据,即观测数据的微小变动可能导致解变动很大,而且在现实生活中,由于观测数据受噪声的污染,方程的解可能偏离真解相当远。因此,无论是理论方面还是实际操作方面,超分辨率和图像去运动模糊都是十分困难的问题。本文回顾了超分辨率和图像去模糊的研究现状,然后提出了自己的视频超分辨率算法和去运动模糊算法。本文提出了一种新的适用压缩域的视频超分辨率方法,该方法将整个超分辨率过程分成两阶段：Step1.利用解压缩低分辨率压缩视频得到的信息,使用贝叶斯框架,来分别得到当前活动窗口内的各初始单帧超分辨率估计；Step2.利用当前窗口内的各单帧超分辨率图像来重建目标帧的最终的超分辨率图像。具体研究成果包括：1.改进了成像模型和压缩模型,削弱了活动窗口内超分辨率重建结果和相邻高分辨率图像间的配准误差对超分辨率结果的影响。2.提出了新的超分辨率重建结果的混合先验模型,它由两部分组成：全局先验模型和局部先验模型。全局先验模型：使超分辨率重建图像的梯度分布尽可能接近解压缩后的低分辨率图像的梯度分布。通过全局先验模型,使超分辨率结果能保持边缘清晰；局部先验模型：如果LR图像的局部区域是平滑的,那么对应的HR图像区域也应该是平滑的。通过局部先验模型抑制了超分辨率结果的“振铃”噪声。3.使用IRLS(迭代重加权最小二乘法)来构建鲁棒性强的目标函数,从而削弱离群点和噪声的影响,减少超分辨率结果包含的人工痕迹。在去运动模糊方面,本文的主要贡献包括：1.提出了新的基于透明度的双图像平面运动估计方法,平面运动包括旋转运动加平移运动。这种方法利用了清晰图像边界的透明度的二值性来减少盲复原问题的病态性。2.减少了用户输入的信息,在模糊核的估计过程中,不需要用户交互。仅仅在得到透明度信息的过程中,需要用户输入部分笔划来表明前景和背景。3.提出了新的运动模糊离散模型和新的平面模糊运动描述符,使最优化算法更加简单。4.改进了经典Richardson-Lucy去卷积算法,使之适合本文提出的新的运动模糊离散模型。5.使用了IRLS(迭代重加权最小二乘法)来构建目标函数,其输入为两幅模糊图像,使去模糊效果更好,更准确。更多还原

【Abstract】 This thesis is concerned with super resolution and image blind restoration, one of the most challenging problems in Computer Vision. Mathematically, both super resolution and image blind restoration are inherently ill-posed inverse problem. Such problems may be tackled by constraining the solution space according to a priori knowledge of the form of the solution.This dissertation first reviews the super resolution approaches and image deblur approaches, then presents a new robust super resolution algorithm of compressed video and a new planar motion deblurring method from dual images using transparency.Before super resolution with multiple frames, we first isolate the frames individually, and get their corresponding initial super resolution estimates within the Bayesian framework by exploiting the information available in the compressed bitstream. Then a final SR image can be obtained with the neighboring initial estimates. In details:1. It formulates the process from HR image to the corresponding decompressed LR frame di-rectely to exclude the register errors between the object hr image and the neighbouring im-ages.2. It presents a new latent image prior, which has two components:the global prior, which leads to edge-preserving SR images, and the local prior, which leads to SR images with less ringing artifacts.3. By using the IRLS approach to get a robust cost function to reject outliers, the SR results with little artifacts have been achieved.In view of motion deblurring, it’s contribution includes:1. It proposes a practical algorithm to deblur rigid-body object undergone planar motion from dual conventional images. The planar motion includes rotation and translation. The blur kernel is spatially-variant and unknown. The ambiguities of deblurring planar motion is reduced by adopting a latent binary transparency map prior as well as the information in two motion blurred images. 2. It presents a new motion blur descriptor to model the general planar motion of a (planar) rigid object, which makes our optimized algorithm to be of low computational complexity.3. It reduces user assistance and no user assistance is needed after the fractional transparency map has been calculated.4. It introduces how to modify the Richardson-Lucy algorithm to incorporate our general planar motion blur descriptor.5. It demonstrates that the planar motion RL algorithm can incorporate an iterative reweight-ed least square (IRLS) approach to improve the deblurred results, with two input images complementary to each other.更多还原