Research On Video And Image Denoising Method Based On Nonlocal Method

Images and videos are important media for humans to observe and understand the world because they can capture and convey real-world information.However,digital images and videos may be affected by interference and noise signals during acquisition,transmission,and storage,which may adversely affect subsequent image and video processing tasks,thereby reducing the accuracy of results.In order to reduce noise in videos and images,existing research works have established various denoising models based on image priors or deep learning.While these models are excellent at removing noise,they do not aim to handle all types of noise but are optimized for specific types of noise.In addition,these models usually have high requirements for training data and computing resources,which also limits their application range.While block-level non-local self-similar priors have been successfully used for video denoising tasks,this limits the performance of non-local self-similar video denoising since the voxels in a single video block may not be completely similar.Most of the existing video denoising methods require complex optical flow estimation,and inaccurate optical flow estimation will also lead to artifacts in the denoised video;and in the case of high noise,the original BM3 D algorithm cannot effectively remove the isolated intensity noise.To solve these problems,this paper proposes the following two denoising algorithms:First,this paper lifts the non-local self-similarity prior of 3-D signals from block-level to voxel-level,and proposes a video denoising method based on voxel-level non-local methods.The method first performs three-dimensional cuboid matching,and then converts each cuboid into a column vector,and then stitches all the column vectors into a two-dimensional voxel matrix.Row matching is performed on a 2D matrix to ensure that the resulting voxel groups are completely self-similar.By performing a simple separable Haar transform on self-similar voxel groups,the noise can be effectively separated from the real signal,and the ideal video denoising effect can be obtained.Since the matching process automatically follows the optical flow trend,this method does not need to estimate the optical flow in advance.In addition,the line matching process automatically fits curved edges in the video,ensuring better preservation of video details in the spatio-temporal domain during denoising.Experimental results on real datasets show that compared with existing deep learning methods,the video denoising method proposed in this paper has an average peak signal-to-noise ratio(PSNR)of 2.0d B higher and an advanced effect on subjective visual quality.Second,this paper proposes a denoising method for BM3 D images using dual hard thresholding.The algorithm uses dual hard thresholding,coefficient hard thresholding and structural hard thresholding.The whole process includes two important steps: basic estimation and final estimation,and each step is further divided into three small steps: block matching,collaborative filtering and aggregation.In the basic estimation stage,the sparsity of the transformed coefficients is further enhanced by reducing the coefficient hard threshold and adding a structural hard threshold in three dimensions to replace the unique coefficient hard threshold in the original BM3 D algorithm.Finally,the improved algorithm was compared with the original BM3 D algorithm on the public benchmark dataset under high noise conditions.Experimental results show that the peak signal-to-noise ratio(PSNR)of the proposed method is 0.15 d B higher than the original BM3 D algorithm.It shows that this method can effectively attenuate and eliminate the noise signal,thereby improving the quality and clarity of the image.