Video denoising method based on voxel-level non-local model

Abstract

The invention discloses a video denoising method based on a voxel-level non-local model. The method comprises first-stage preliminary denoising and second-stage fine denoising. The method has the advantages that all blocks matched with three-dimensional blocks are scanned into column vectors, then row matching is carried out, and the most similar voxel group is obtained. Compared with an existingnon-local method based on two-dimensional image blocks, the method for executing image denoising on the similar voxel group has multiple advantages that firstly, cross-frame long and narrow signals cannot be regarded as noise when three-dimensional block matching is executed; 2, detail information in the video can be better reserved while noise is removed; and thirdly, Gaussian noise and signal-related noise in the video can be effectively removed.

Description

technical field [0001] The invention relates to the technical field of video processing, in particular to a video denoising method of similar voxel matching Haar transform. Background technique [0002] Even high-end cameras shoot videos under low-light conditions with more or less noise, and high-speed cameras using shorter exposure times will introduce stronger noise. In addition, cheap, low-quality sensors are still widely used in mobile phones or surveillance cameras, and such low-end camera equipment will inevitably introduce noise even in videos captured in well-lit environments, especially in difficult shooting conditions Videos obtained in low light conditions (low light, small sensors, etc.) contain more noise. Therefore, video denoising remains an important step in video processing and underlying machine vision research. [0003] The existing research literature on video denoising is much less than that on image denoising. Most of the current video denoising met...

AI Technical Summary

Problems solved by technology

Most of the current video denoising methods are based on two-dimensional image block-level self-similarity, the popular video denoising method VBM3D [1] It is to perform two-dimensional block matching between frames, but some narrow and long signals in a video frame are often regarded as noise in a certain frame, and the video denoising method based on two-dimensional image blocks can easily be regarded as noise and was removed

VBM4D [2] Although the method performs the matching of three-dimensional cubes and a separable four-dimensional transformation, the first three-dimensional transformation is similar to the separable three-dimensional transformation in [1]. Although the video denoising results are generally improved compared with the VBM3D method, But it is still unsatisfactory, especially in the denoising process, the video details cannot be well preserved

[0004] In recent years, although the video denoising problem has been developed rapidly, the video denoising method based on two-dimensional block-level self-similarity seems powerless to signal-related noise in real video; although the method based on deep learning has been obtained in recent years Rapid development, but the method based on deep learning relies too much on the training data set. Generally, when training and testing on the same data set, the test results are ideal, but when switching to another data set for testing, the results are very poor

Images