CASIA OpenIR  > 毕业生  > 博士学位论文
基于耦合正则化的图像去噪与超分辨率重建算法研究
Alternative TitleResearch on Image Denoising and Super-resolution Reconstruction Based on Coupled Regularization
李林
Subtype工学博士
Thesis Advisor张文生
2014-05-28
Degree Grantor中国科学院大学
Place of Conferral中国科学院自动化研究所
Degree Discipline计算机应用技术
Keyword核回归 超分辨率 Bregman迭代 全变分 稀疏表示 Kernel Regression Super-resolution Bregman Iteration Total Variation Sparse Representation
Abstract人类获取的信息中有高达85%来自图像,我国大中城市的视频监控逐步实现了全覆盖。受成像设备、成像环境、传输带宽以及传输编解码因素的影响,采集的图像经常存在模糊、噪声和空间分辨率低的现象。为了在不改变采集和传输条件的情况下获取高质量的图像,图像去噪和超分辨率重建应运而生。然而,图像去噪和超分辨率重建问题是病态、不适定问题,由于缺少深入的理论基础和有效的技术支撑,图像去噪和超分辨率重建一直是图像处理领域的难点。将图像的先验知识与正则化技术融合,进而解决图像去噪和超分辨率重建难题,是该领域研究的热点。本文选择基于耦合正则化的图像去噪与超分辨率重建算法研究,具有重要的理论意义和广泛的应用前景。 在图像去噪研究中,提出了一种基于可控核回归(steering kernel regression)全变分的图像去噪算法,保证了去噪过程中图像的边缘和结构信息。在图像超分辨率重建研究中,提出了一种基于组合全变分的超分辨率重建算法,充分利用图像的局部和非局部信息,克服了传统的基于重建的算法平滑图像纹理的不足;提出了一种基于稀疏表示残差的超分辨率重建算法,在全变分模型的基础上,加入稀疏表示先验知识学习图像的细节信息,改善了图像超分辨率重建的性能。 本文主要工作与贡献如下: 1. 针对传统的核回归去噪算法模糊图像边缘信息的问题,提出了一种基于可控核回归全变分(steering kernel regression total variation)的图像去噪算法。该算法综合基于核回归的算法和全变分算法的优点,将核回归算法扩展到有界变分空间,并使用可控核回归全变分项分析辐射度差异(radiometric difference)获得图像变分域的局部结构属性;采用非局部的核回归全变分学习图像的非局部相似信息,并将两者结合建立了基于核回归全变分的图像去噪模型;使用Split Bregman 迭代方法对提出的算法进行优化求解。通过与主流算法在基准数据集上的实验对比,发现本文提出的算法在有效去除噪声的同时具有保持图像结构和细节信息的能力。 2. 针对基于非局部全变分的超分辨率重建算法无法保持图像局部结构的问题,提出了一种基于组合全变分(Combined Total Variation,CTV)的图像超分辨率算法。该算法使用非局部全变分项(Nonlocal Total Variation)学习图像非局部块之间的相似性,将其作为超分辨率重建的非局部先验约束;使用提出的可控核回归项全变分(SKRTV)项学习图像的局部结构属性,将其作为超分辨率重建的局部先验约束;通过融合局部和非局部的先验知识,协调约束保真项与两个先验正则化项,建立了基于最大后验概率(MAP)的图像超分辨率模型。通过与国际先进算法的使用对比,发现本文提出的算法在保持清晰的边缘的同时可以获得比较丰富的细节。 3. 针对基于稀疏表示的超分辨率重建算法过于依赖训练集的问题,提出了一种基于稀疏表示残差(Sparse Representation Residual,SRR)的图像超分辨率重建算法。该算法将基于重建的算法与基于学习的算法进行了桥接,并通过融合全局保真项、稀疏表示残差项、稀疏表示项与全变分正则化项,建立了超分辨率重建的数学模型;交替迭代基...
Other AbstractMore than 85% of information is accessed by images, and video monitoring has gradually achieved full coverage in Chinese large and medium-sized city. Fuzzy, noise and low spatial resolution phenomenon is existed by the influence of imaging device, imaging environment, transmission bandwidth and transmission decoding. In order obtain high quality images, we need to use the image denosing and super resolution(SR) technique under given qcquisition and transmission conditions. However, image denoising and super resolution is ill-posed problems, because of lacking theoretical basis and effective technical support, image denoising and super resolution is very different. By using regularization-based method to solve image denoising and super resolution problem is a hot topic in image processing area. Coupled-regularization-based image denoising and super resolution method is chosen in this thesis, which has important theoretical significance and wide application prospect. During the research of image denoising method, we proposed a steering kernel regression total variation based denoising method, which could preserve the edge and structure information. During the research of image super resolution method, we proposed a combined total variation based SR method, which could overcome the shortcoming of texture smoothing of reconstruction-based method. The main content of this paper is as follows: 1. In order to overcome the disadvantage of kernel regression-based image denoising method, which usually blur edge, this thesis proposes a denoising algorithm based on steering kernel regression total variation. This algorithm obtains the local structure through analyzing radiometric difference, then it estimates the weight using kernel function through structure to get the reconstruction image. The Split Bregman iteration algorithm is applied to solve the proposed model. Experiments show that this method can both remove the noise and retain the structure of images. 2. In order to overcome the disadvantage of nonlocal total variation-based image super resolution method, which could not reserve the structure of images, this thesis comes up with a super resolution algorithm based on combined total variation. Utilizing redundancy of natural image patches, this algorithm learns the prior information of nonlocal similarity by adding Nonlocal Total Variation. By adding the nonlocal priori constraint to the super-resolution model, it can make good use of these nonlocal i...
shelfnumXWLW2034
Other Identifier201118014629082
Language中文
Document Type学位论文
Identifierhttp://ir.ia.ac.cn/handle/173211/6630
Collection毕业生_博士学位论文
Recommended Citation
GB/T 7714
李林. 基于耦合正则化的图像去噪与超分辨率重建算法研究[D]. 中国科学院自动化研究所. 中国科学院大学,2014.
Files in This Item:
File Name/Size DocType Version Access License
CASIA_20111801462908(20516KB) 暂不开放CC BY-NC-SAApplication Full Text
Related Services
Recommend this item
Bookmark
Usage statistics
Export to Endnote
Google Scholar
Similar articles in Google Scholar
[李林]'s Articles
Baidu academic
Similar articles in Baidu academic
[李林]'s Articles
Bing Scholar
Similar articles in Bing Scholar
[李林]'s Articles
Terms of Use
No data!
Social Bookmark/Share
All comments (0)
No comment.
 

Items in the repository are protected by copyright, with all rights reserved, unless otherwise indicated.