【作者】 刘松；

【导师】 朱路；

【作者基本信息】 华东交通大学 ， 信息与通信工程， 2019， 硕士

【摘要】 在人类日常生活中,图像扮演了越来越重要的角色,图像蕴含了丰富的信息,但图像受到各种各样的因素的影响,从而导致其质量不佳。一般情况下,图像会被噪声污染,直接影响到图像的视觉效果,为了解决该问题,图像去噪技术因此得到了成熟的发展,主要目的为既能高效地去除图像噪声,又能有效地保留图像的有用信息,包括图像的纹理和边缘信息等。近些年来,随着信号的稀疏表示理论的不断完善,基于稀疏表示理论的字典学习方法得到了成熟的发展。考虑到原始图像在合适的字典下具有稀疏性而噪声通常无这一特性,因此,研究学者们为解决图像去噪问题提出了很多有效解决方案。基于稀疏表示的图像去噪方法对图像噪声方差难以确定,且传统字典学习方法难以解决参数自动选择问题。而非参数贝叶斯字典学习方法能有效地解决该问题。考虑到图像在梯度域具有良好稀疏性以及非局部自相似特性,分别提出了稀疏梯度域非参数贝叶斯字典学习图像去噪方法和梯度域补丁分组贝叶斯学习图像去噪方法。图像在梯度域的稀疏性一般优于空间域,提出稀疏梯度域非参数贝叶斯字典学习图像去噪方法。考虑整个图像去噪模型是一个多变量耦合问题,难以直接求解。利用Bregman和交替迭代方法把该问题分解为若干个子问题,再利用最小二乘法和非参数贝叶斯字典学习BPFA(Beta Process Factor Analysis)方法求解这些子问题。非参数贝叶斯字典学习模型复杂,难以直接求解模型参数;采用吉布斯(Gibbs)采样依次交替迭代求解模型参数,从而我们能获得最优字典和稀疏表示。整个算法相较于GradDLRec算法,具有良好的去噪性能。考虑图像的结构信息,采用聚类方法对图像块进行分类处理,提出梯度域图像补丁分组贝叶斯学习图像去噪方法。该方法依据图像的非局部自相似特性,先对图像补丁进行分组,然后利用了高斯混合模型对梯度图像的分组补丁进行匹配,得到不同的补丁组分,最后利用BPFA字典学习方法重建各个组分潜在的干净的梯度图像结构。在BPFA字典学习模型上施加结构聚类,整个图像去噪模型得到很大改善。更多还原

【Abstract】 In human daily life,image plays an increasingly important role.Image contains abundant information,but image is affected by various factors,resulting in poor quality.Generally,the image is polluted by noise,which directly affects the visual effect of the image.In order to solve this problem,image denoising technology has been developed maturely.The main purpose is not only to effectively remove image noise,but also to effectively retain useful information of the image,including texture and edge information of the image.In recent years,with the continuous improvement of signal sparse representation theory,the dictionary learning method based on sparse representation theory has been developed maturely.Considering that the original image is sparse in a proper dictionary and the noise usually does not have this characteristic,researchers have put forward many effective solutions to solve the problem of image denoising.The image denoising method based on sparse representation is difficult to determine the variance of image noise,and the traditional dictionary learning method is difficult to solve the problem of automatic parameter selection.The nonparametric Bayesian dictionary learning method can effectively solve this problem.Considering that the image has good sparsity and nonlocal self-similarity in gradient domain,the nonparametric Bayesian dictionary learning image denoising method in sparse gradient domain and the patch grouping Bayesian learning image denoising method in gradient domain are proposed respectively.The sparsity of image in gradient domain is generally better than that in spatial domain.A nonparametric Bayesian dictionary learning method for image denoising in sparse gradient domain is proposed.Considering that the whole image denoising model is a multi-variable coupling problem,it is difficult to solve directly.Bregman method and alternating iteration method are used to decompose the problem into several sub-problems,and then the least squares method and nonparametric Bayesian dictionary learning BPFA(Beta Process Factor Analysis)method are used to solve these sub-problems.The nonparametric Bayesian dictionary learning model is complex and difficult to directly solve the model parameters.Gibbs sampling is usedto iterate the model parameters in turn and we can obtain the optimal dictionary and sparse representation.Compared with GradDLRec algorithm,the whole algorithm has good denoising performance.Considering the structure information of the image,the image blocks are classified by clustering method,and the gradient domain image patch grouping Bayesian learning image denoising method is proposed.According to the nonlocal self-similarity of images,firstly,the method groups the image patches.Then we use the Gaussian mixture model to match the grouped patches of gradient images and get different patch components.Finally,the potential clean gradient image structure of each component is reconstructed by BPFA dictionary learning method.Structural clustering is applied to the BPFA dictionary learning model,and the whole image denoising model is greatly improved.更多还原