基于稀疏表示理论的图像去噪算法研究

发布时间：2018-05-20 18:16

本文选题：图像去噪 + 稀疏表示　；参考：《昆明理工大学》2017年硕士论文

【摘要】：随着数字化多媒体的发展,人们对于图像质量的要求也逐渐变高,然而图像噪声污染很大程度上降低了图像的质量,因此图像去噪是图像处理的研究重点。传统的图像去噪方法虽然实现较为容易,但却容易丢失图像结构纹理信息。近年来随着压缩感知的推广和兴起,使得稀疏表示理论得到了很大的关注。本文主要对基于稀疏表示理论的图像去噪算法展开研究,对传统稀疏表示去噪算法进行改进,具体工作有以下三方面:第一,传统的图像去噪方法虽然容易实现,但容易丢失图像结构纹理信息,针对传统去噪方法存在的不足,本文提出一种基于形态学成分分析(Morphological Component Analysis,MCA)将图像拆分为纹理部分和结构部分,对纹理部分采用基于全相位离散余弦双正交变换(All Phase Discrete Cosine Biorthogonal Transform,APDCBT)字典的稀疏表示算法去噪,对结构部分采用三维块匹配BM3D(Block-Matching and 3-D Filtering)算法去噪。最后将两部分合起来得到最终去噪图像。第二,针对传统K奇异值分析(K-SVD)的图像去噪方法对于信号利用率不足,本文采用稀疏贝叶斯学习(Sparse Bayesian Learning,SBL)对图像信号进行预处理。因为噪声原子存在于字典更新之后得到的字典中,所以结合Bartlett检验法将噪声原子裁剪掉。第三,针对传统K奇异值分析(K-SVD)的图像去噪方法运行时间较长以及凸优化算法中最速下降算法的锯齿现象的缺点,本文提出了一种改进的最速下降算法与正交匹配追踪(orthogonalmatchingpursuit,OMP)相结合的方法,有效解决了运行时间和锯齿现象。本文通过对比试验,仿真结果表明,本文提出的去噪方法相对于传统去噪方法更好的保留图像的细节特征和结构信息,获得更高的峰值信噪比值(PSNR)。相比于传统K-SVD图像去噪方法本文方法能更好地滤除噪声,运行时间得以提高。为了更好的反应本文方法取得较好的实验结果,将本文方法做成Matlab GUI更直观的对比显示。
[Abstract]:With the development of digital multimedia, the demand for image quality is becoming higher and higher. However, image noise pollution greatly reduces the image quality, so image denoising is the focus of image processing. Although the traditional image denoising method is easy to realize, it is easy to lose the texture information of image structure. In recent years, with the expansion and rise of compressed perception, sparse representation theory has been paid much attention. In this paper, the image denoising algorithm based on sparse representation theory is studied, and the traditional sparse representation denoising algorithm is improved. The specific work is as follows: first, the traditional image denoising method is easy to realize. But it is easy to lose the texture information of image structure. Aiming at the shortcomings of the traditional denoising methods, this paper proposes a new method based on morphological component analysis (MCA) to divide the image into texture and structure parts. The sparse representation algorithm based on all Phase Discrete Cosine Biorthogonal transform (all Phase Discrete Cosine Biorthogonal transform) dictionary is used for texture part, and the 3D block matching BM3D(Block-Matching and 3-D filtering algorithm is used for the structure part. Finally, the two parts are combined to get the final denoising image. Secondly, because the traditional K-SVD (K-SVD) image denoising method is insufficient for signal utilization, sparse Bayesian learning Bayesian learning (SBL) is used to preprocess the image signal. Because noise atoms exist in the dictionary after the dictionary is updated, the noise atoms are clipped by Bartlett test. Thirdly, aiming at the disadvantages of the traditional K-SVD (K-SVD) image denoising method and the sawtooth phenomenon of the steepest descent algorithm in the convex optimization algorithm, In this paper, an improved steepest descent algorithm combined with orthogonal matching tracking algorithm (OMP) is proposed, which solves the problem of running time and sawtooth effectively. The simulation results show that compared with the traditional denoising method, the proposed denoising method preserves the detailed features and structure information of the image better, and obtains a higher PSNRN than the traditional denoising method. Compared with the traditional K-SVD image denoising method, this method can filter noise better and improve the running time. In order to better reflect the experimental results of this method, this method is made into Matlab GUI more intuitionistic contrast display.
【学位授予单位】：昆明理工大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41

【参考文献】