【作者】 陈晓钢；

【导师】 杨杰；

【作者基本信息】 上海交通大学 ， 模式识别与智能系统， 2013， 博士

【摘要】 图像恢复是计算机视觉领域的一个重要研究课题。拍摄环境中光照条件不充分时，相机往往需要较长的曝光时间使图像充分曝光，而长时间曝光则易引起图像运动模糊。另一方面，相机可以通过设置高感光度，提升传感器的光线敏感度来减少曝光时间并避免图像模糊。然而，高感光度会放大图像噪声，使图像噪点突出并影响图像质量。图像模糊和图像噪声干扰均可看作清晰图像的退化。本文对这两类图像退化进行研究，并提出有效的图像恢复算法。对于图像直线运动模糊，本文分析了直线运动模糊在空域和频域的方向性高频能量分布特性，提出了基于方向性高通滤波的模糊特征分析框架。基于提出的方向性高通滤波，本文方法可有效判别图像模糊区域，估计图像模糊方向，较传统模糊方向估计方法更为简洁并且准确。对于更普遍的非直线运动模糊，本文分析了自然场景图像和文本图像梯度分布统计特征的差异性。根据图像统计特征，本文提出了基于梯度域的图像盲反卷积以及结合图像局部上下界约束的文本图像恢复算法。实验显示本文提出的文本图像去模糊实现了比同类算法更优的恢复效果。同时，本文进一步分析了图像去模糊问题假设模型的可靠性。对于空不变模糊，传统图像去模糊方法往往假定模糊图像由清晰图像和点扩散函数线性卷积得到，然而实际情况下相机响应函数在图像生成过程中会使模糊图像偏离卷积模型。因此线性卷积模型应用于实际图像去模糊时，存在模型近似误差，基于该假设条件的图像反卷积也会引起图像振铃。本文探讨了相机响应函数对图像反卷积的影响，尤其是对图像低频和高频区域的不同扰动，并理论分析了卷积模型引入的误差和误差界。此外，本文根据清晰图像和模糊图像的灰度分布特征提出了相机响应函数估计算法。图像去噪方面，非局部均值和BM3D等基于图像块相似性的图像去噪是目前比较流行的算法。这类算法能有效滤除图像噪声，但由于图像块相似性匹配计算量过大，实时运算很困难。针对计算量大的缺点，本文提出基于图像块测地线距离的图像去噪算法。通过研究图像块测地线距离最短路径的性质，本文提出图像块距离的快速近似计算，并采用多尺度算法实现图像去噪。本文提出的算法与非局部均值算法性能相当，但计算效率远高于该算法以及其他基于图像块相似度的图像去噪算法。更多还原

【Abstract】 Image restoration has been a long-standing problem in computer vision. Under low-lightconditions, a camera requires either long exposure or a high ISO setting in order to obtain abright image. However, long exposure will result in image motion blur, and high ISO will amplifynoises in the captured image. These two situations can be regarded as image degradation. In thispaper, image restoration methods are proposed to estimate sharp and clean image from acaptured blurry or noisy image.For image motion blur, the properties of linear motion blur in spatial and frequency domainare investigated. The directional high-pass filter is proposed to identify motion direction anddetect motion blurred regions. Specifically, a closed-form solution for motion directionestimation is derived. Regarding general motion blur, this paper studied the gradient statistics ofnatural-scene images and document images. Gradient domain blind deconvolution is proposedand content-aware document image deblurring is presented. Image local constraints areconsidered in order to suppress ringing artifacts. The results produced by the proposed methodsare comparable to that of the state-of-the-art methods.A common assumption in image deblurring is that a blurry image is a linear convolution resultof a sharp image by a Point Spread Function. Nevertheless, in a real camera system, theconvolution is carried out in irradiance domain and Camera Response Function (CRF) willnonlinearly map the convolved irradiance to the output intensity image. This paper presents acomprehensive study on the effects of CRFs on motion deblurring. The approximation errorcaused by the direct intensity convolution is analyzed. We prove that the intensity-basedconvolution closely approximates the irradiance model at low frequency regions, but it will havelarge deviation at high frequency regions. Then, we further propose a CRF estimation methodbased on a pair of sharp/blurred images.Finally, image denoising problem is reviewed. Patch-based denoising, like Non-Local Means(NLM) and BM3D, is able to produce high-quality result. However, they require expensivepair-wise patch comparisons. This paper introduces a Patch Geodesic (PG) distance metric forpatch comparison. In order to reduce noise at multiple scales, we adopt Laplacian-Gaussianimage decomposition and apply PG-based denoising at each scale. It achieves comparable qualityas the state-of-the-art methods. Since PG path can be efficiently approximated by minimum hoppath, the proposed method is a few orders of magnitude faster than traditional patch-baseddenoising methods without losing quality.更多还原