Image Denoising Algorithm Based On Special GMM Prior And Rotational Prior

Image signals play an important role in modern communication technology.However,interference is everywhere in the real world,and the image is distorted in the process of acquisition and dissemination due to the change of equipment or environmental conditions.The research shows that the noise formed by a large number of interfering signals obeys a certain distribution,so it becomes a hot topic to use these laws to pre-process the image.In recent years,the image de-noise model based on non-local self-similarity prior(Nonlocal self-similarity,NSS)has received widespread attention because of the repeated structure of natural image patches.Similar patches collected by NSS prior are sparse,which can be used to estimate potential low-rank subspaces.On the other hand,the modeling of natural images,such as Gaussian mixture models(GMM),has been successful in all aspects of computer vision.However,the existing denoising algorithms based on the GMM model have the following two problems: 1.How to train a GMM model that processes specific types of images.2.How to accelerate the matching of rotational similar patches by GMM model.This thesis proposes solutions to the two problems of external priors.The main contributions of this thesis are as follows:(1)We first explain the probabilistic meaning of several external denoising algorithms from the perspective of Bayesian theory,and analyze the applicability of GMM and the possible underfitting problems for large data sets.Subsequently,the concept of dedicated GMM was proposed,and a dedicated GMM model was trained for a specific data set to adapt to the denoising problem of similar images.This training method classifies the data before training and effectively reduces the underfitting phenomenon of GMM caused by the large data set.Experiments show that our method effectively reduces the number of samples required to train GMM model,and the indexes are better than the existing GMM-based denoising method in database image denoising.(2)Aiming at the shortcomings of the previous matching strategy,we proposed a new rotation clustering method.Foremost,the noisy image is pre-rotated as a whole to form a set,and then the GMM model will simultaneously guide the clustering of each element in the set.The parallel computing framework we designed only needs to prerotate the entire image to avoid rotating the patches dynamically during the denoising process,so it has a high computational efficiency.The intra-cluster merger strategy is implemented to match similar patches and their rotational versions,after the parallel clustering is over.The experiment result shows that compared with the traditional image denoising algorithm,this method can improve the denoising performance by using the rotational prior.Our PSNR index increased by up to 0.35 d B on images with strong geometric rotation characteristics,such as the Checkerboard image.Combining the complexity of computation and the increment of denoising index,our algorithm has a strong advantage.In the experiment,we have realized the acceleration program of the denoising algorithm designed by multi-core CPU,and the experiment shows that our algorithm can obviously reduce the speed of operation by multi-core acceleration.