Research On Image Denosing Based On Sparse Representation

In the field of digital image processing, the research on image denoising has been one of the most important aspects. In the same time, sparse representation theory has raised a research upsurge in recent years. Image sparse representation can extract nature feature to express the image in a simple way. Based on the excellent characteristics, signal sparse expression can be applied to many fields of signal processing. Traditional image denoising approaches project the image signal onto the transform domain so the signal and the noise can be separated, but in the transform domain, the noise cannot be separated completely with the signal, so in the process of image denoising, the image information is also damaged. However, the image denoising algorithm based on sparse representation can separate the noise completely with the signal, because the noise does not have sparse component. The K-SVD image denoising algorithm is studied firstly. Although K-SVD algorithm is adaptive, and it can remove the noise from the image excellently, there is still residual noise after denoising, in the same time, when the image is damaged very seriously, the performance is not so obvious. Based on this, this thesis puts forward a double image denosing algorithm. The effect of this algorithm is improved obviously compared to K-SVD algorithm. The main research contents of this thesis include the following aspects:First, the basic theory of sparse representation and related sparse decomposition algorithm is introduced. And then the description of the image denoising model based on the theory of sparse representation comes after.Second, several algorithms of image denoising have been researched in this thesis, mainly focusing on the image denoising algorithm based on K-SVD, and through simulation analysis, the results show that K-SVD algorithm has obvious advantages compares with the traditional algorithm.Third, how to construct the dictionary has been studied, it is one of the core problems of K-SVD image denosing. There are two general kinds of dictionaries:analysis dictionary and learning dictionary, both of them have two sides. Through the study on the advantages of these two kinds of dictionaries, a cascade dictionary is constructed; this cascade dictionary can not only have self-changing adaptation but also maintain the structure characteristics of the dictionary. The simulation results have been showed that the cascade dictionary can restore the image texture information better.At last, in order to remove noises effectively, a double image denosing algorithm has been proposed in this thesis. Firstly, the noise from signal is removed by K-SVD algorithm and then the output the denoised image is denoised again using another iterative denoising algorithm. Because the noise has been carried out by K-SVD firstly, the number of iterations during the second denoising can be reduced, and the computational complexity of image denoising can be reduced as well. The simulation experiment shows that the denoising effect of the proposed algorithm is more obvious than K-SVD algorithm, and the more image noise, the better denoising performance.