基于稀疏表示的图像去噪算法
发布时间：2018-09-08 14:46
【摘要】：图像处理的最基本问题是图像去噪，随着压缩感知的兴起与推广，越来越多的学者开始关注稀疏表示理论及其应用，基于稀疏表示的图像去噪成为近年来该领域的前沿研究课题。论文研究了稀疏表示理论，探讨了原子库构造和稀疏分解两个关键性问题。在此基础上，，对基于过完备字典稀疏表示的图像去噪方法进行了探索性研究。 在稀疏表示理论中，字典的构造方法有两种：一种是选取固定基组构成解析型的字典；另一种是基于训练样本学习得到自适应字典。固定基构造的字典虽然不能自适应表示信号的结构特征，但具有很快的执行速度，在实际中仍然广泛应用。论文选择课题组前期提出的全相位双正交变换(APBT)构造原子库，并将几种基函数组合成混合原子库，提出基于该类字典表示的图像去噪方法，取得了较好的实验结果。 基于学习方法构造的冗余字典可更加准确地提取信号的结构特征，也是近几年的研究热点。论文在研究了基于KSVD字典学习的图像去噪算法的基础上，将相关系数匹配准则和字典裁剪方法相结合，提出一种改进的字典学习算法，进一步，为了利用图像的非局部自相似性信息，提出将自相似性作为一个约束正则项融入到图像去噪模型，提出基于改进字典学习和非局部自相似性的图像去噪算法。大量实验验证，与传统KSVD去噪方法相比，该方法在提高同质区域平滑性的同时还能保留更多的纹理、边缘等细节特征。
[Abstract]:The basic problem of image processing is image denoising. With the rise and popularization of compressed sensing, more and more scholars begin to pay attention to sparse representation theory and its application. Image denoising based on sparse representation has become a frontier research topic in this field in recent years. On this basis, an exploratory study of image denoising method based on sparse representation of over-complete dictionary is carried out.
In sparse representation theory, there are two ways to construct a dictionary: one is to select a fixed base group to form an analytic dictionary; the other is to learn an adaptive dictionary based on training samples. This paper chooses all-phase biorthogonal transform (APBT) proposed by our research group to construct the atom library, and combines several basis functions into a hybrid atom library.
Redundant dictionaries based on learning methods can extract the structural features of signals more accurately, which is also a research hotspot in recent years. On the basis of studying the image denoising algorithm based on KSVD dictionary learning, this paper proposes an improved dictionary learning algorithm by combining the correlation coefficient matching criterion with the dictionary clipping method. In order to utilize the non-local self-similarity information of images, a new image denoising algorithm based on improved dictionary learning and non-local self-similarity is proposed, which incorporates self-similarity as a constraint regular term into the image denoising model. It can retain more details such as texture, edge and so on.
【学位授予单位】：天津大学
【学位级别】：硕士
【学位授予年份】：2014
【分类号】：TP391.41
本文编号：2230833
[Abstract]:The basic problem of image processing is image denoising. With the rise and popularization of compressed sensing, more and more scholars begin to pay attention to sparse representation theory and its application. Image denoising based on sparse representation has become a frontier research topic in this field in recent years. On this basis, an exploratory study of image denoising method based on sparse representation of over-complete dictionary is carried out.
In sparse representation theory, there are two ways to construct a dictionary: one is to select a fixed base group to form an analytic dictionary; the other is to learn an adaptive dictionary based on training samples. This paper chooses all-phase biorthogonal transform (APBT) proposed by our research group to construct the atom library, and combines several basis functions into a hybrid atom library.
Redundant dictionaries based on learning methods can extract the structural features of signals more accurately, which is also a research hotspot in recent years. On the basis of studying the image denoising algorithm based on KSVD dictionary learning, this paper proposes an improved dictionary learning algorithm by combining the correlation coefficient matching criterion with the dictionary clipping method. In order to utilize the non-local self-similarity information of images, a new image denoising algorithm based on improved dictionary learning and non-local self-similarity is proposed, which incorporates self-similarity as a constraint regular term into the image denoising model. It can retain more details such as texture, edge and so on.
【学位授予单位】：天津大学
【学位级别】：硕士
【学位授予年份】：2014
【分类号】：TP391.41
【参考文献】
相关期刊论文 前7条
1 蔡泽民;赖剑煌;;一种基于超完备字典学习的图像去噪方法[J];电子学报;2009年02期
2 王刚;王娟;王德华;张志峰;肖亮;贺安之;;基于Contourlet变换域统计模型的路面图像去噪算法[J];光电子.激光;2009年10期
3 周季夫;钟诚文;尹世群;解建飞;张勇;;基于GPGPU的Lattice-Boltzmann数值模拟算法[J];计算机辅助设计与图形学学报;2008年07期
4 刘琚,何振亚;利用SVD对带噪声的模糊图像进行盲复原[J];数据采集与处理;2002年01期
5 侯正信;离散余弦列率滤波器的设计及应用[J];天津大学学报;1999年03期
6 尹忠科,王建英,邵君;基于原子库结构特性的信号稀疏分解[J];西南交通大学学报;2005年02期
7 程文波;王华军;;信号稀疏表示的研究及应用[J];西南石油大学学报(自然科学版);2008年05期
相关博士学位论文 前2条
1 孙玉宝;图像稀疏表示模型及其在图像处理反问题中的应用[D];南京理工大学;2010年
2 王成优;全相位双正交变换理论及其在图像编码中的应用研究[D];天津大学;2010年
本文编号：2230833
本文链接：https://www.wllwen.com/kejilunwen/ruanjiangongchenglunwen/2230833.html
