CASIA OpenIR  > 毕业生  > 博士学位论文
小波域图象超分辨率重构算法研究
其他题名Wavelet-Domain Algorithms For Image Superresolution
赵书斌
学位类型工学博士
导师彭思龙
2003-08-13
学位授予单位中国科学院研究生院
学位授予地点中国科学院自动化研究所
学位专业模式识别与智能系统
关键词小波 隐马尔可夫树模型 超分辨率 规整化 Wavelets Hidden Markov Tree Image Superresolution Regularize
摘要所谓图像超分辨率重构,是指由一幅或几幅低分辨率图像获得一幅高分辨率 图像的过程。一般地,低分辨率图像可以看成高分辨率图像经过模糊和亚采样后 的结果,因而图像的超分辨率重构是一个典型的病态问题。因此,由低分辨率图 像完全重构原来的高分辨率图像是不可能的。 目前,处理图像超分辨率问题的方法可以分为两类。一类是引入关于原高分 辨率的先验知识,然后在该先验知识确定的子空间内寻求满足图像退化过程的一 个解作为对原高分辨率图像的一个估计。另一类是图像类比,即由已知的高分辨 率图像和相应的低分辨率图像通过训练获得低分辨率图像到高分辨率图像的映 射关系,然后将此映射关系应用于低分辨率图像即可获得所需要的高分辨率图 像。第一类方法的关键在于引入先验知识。最初采用的关于自然图像的先验知识 主要是各种光滑性假定,即认为自然图像满足分片线性,一阶光滑性,二阶光滑 性以及一阶和二阶光滑性等条件。后来进行的改进包括线过程,半二次,全变差 和小波域广义高斯分布等。其共同缺点是不能有效区分图像中的平滑区和边缘,并导致边缘的严重模糊。 为克服这些缺点,采用小波域隐马尔可夫树模型作为自然图像的先验模型。 通过不同尺度小波系数的隐状态之间的马尔可夫依赖性刻划小波系数沿尺度的 传递特性并采用混合高斯分布逼近小波系数的边缘分布,小波域隐马尔可夫树模 型准确刻划了自然图像的小波变换的统计特性。通过引入小波域隐马尔可夫树模 型,图像超分辨率问题转化为一个优化问题。小波系数的隐状态所具有的多尺度 图像边缘检测的能力使得我们可以对平滑区和边缘处小波系数进行不同处理,因 而可以有效地保持边缘。 鉴于彩色图像的应用越来越广泛,文中也讨论了彩色图像的超分辨率重构问 题。由于彩色图像三个通道之间存在着相关性,彩色图像超分辨率重构并不是灰 度图像超分辨率重构算法的简单推广。通过三个通道的数据融合得到一幅灰度图 像,然后用该灰度图像协调彩色图像三个通道的超分辨率重构。这种方法较好地 解决了色彩失真问题。 此外,为解决退化过程未知情况下的超分辨率问题,利用多尺度边缘的自相 似特性提出了一种基于小波系数预测的盲图像超分辨率算法。该算法重构出的高 分辨率图像具有较高的信噪比和较好的视觉效果,而同时计算复杂度确很小。 本文提出图像超分辨率重构算法较好地解决了高分辨率图像的边缘模糊问 题,在抑
其他摘要Image superresolution is referred to as generating a high-resolution image from one or several low-resolution images. Generally speaking, a low-resolution image can be considered as the degraded version of the high-resolution image and the degradations are characterized by blurring and downsampling. Like image restoration, image superresolution is an ill-conditioned problem. So extracting the original high-resolution image from low-resolution images is impossible. The only thing that can be done is to find a high-resolution image which is approximate to the original image in some sense. Now, there are two types of algorithms for image superresolution: one is based reconstruction and the other is based learning. According to the reconstruction-based approach, the prior knowledge about the original high-resolution image is introduced and combined with the degradation model to search an estimate of the original high-resolution image. The learning-based method can also be called as image analogies. First, the relations between high-resolution images and the corresponding degraded low-resolution images is learned from training data; then, apply this learned relations to the low-resolution images and obtain the high-resolution images. Obviously, the key of the former approach lies in the introduction of prior knowledge. Traditionally, natural images are assumed to have smoothness of different order, for example, of one or two order. The improved image models include introduction of line process, the half-quadratic model, total variation and wavelet-domain generalized Gaussian distribution. Unfortunately, there is one thing in common to all these prior models: the reconstructed high-resolution images are severely blurred, especially around edges. To overcome this drawback, the wavelet-domain hidden Markov tree-structured (HMT) model is adopted as the prior model of natural images. Modeling the wavelet coefficients of natural images ,by the Markov dependencies of the hidden stare variables and Gaussian mixtures, wavelet-domain HMT precisely characterizes the: statistics of wavelet transform of natural images.By introducing wavelet-domain HMT, image superresolution is converted into a constrained or unconstrained optimization problem. The state probabilities of wavelet coefficients can detect the multiscale edges in images, so different regions of an image can be processed adaptively by partitioning the image into smooth regions and edges and thus edges are properly preserved. Because color images are widely used in many applications, color image superresolution is discussed in this paper. Due to the correlations among three channels of color images, any channel-by-channel superresolution method will result in color distortions. In this paper, a data-fusion-based algorithm is proposed. According to this approach, a gray-scale image is obtained by fusing three channels and then coordinate the superresolution rec
馆藏号XWLW847
其他标识符847
语种中文
文献类型学位论文
条目标识符http://ir.ia.ac.cn/handle/173211/5784
专题毕业生_博士学位论文
推荐引用方式
GB/T 7714
赵书斌. 小波域图象超分辨率重构算法研究[D]. 中国科学院自动化研究所. 中国科学院研究生院,2003.
条目包含的文件
条目无相关文件。
个性服务
推荐该条目
保存到收藏夹
查看访问统计
导出为Endnote文件
谷歌学术
谷歌学术中相似的文章
[赵书斌]的文章
百度学术
百度学术中相似的文章
[赵书斌]的文章
必应学术
必应学术中相似的文章
[赵书斌]的文章
相关权益政策
暂无数据
收藏/分享
所有评论 (0)
暂无评论
 

除非特别说明,本系统中所有内容都受版权保护,并保留所有权利。