Bayer domain image noise reduction system and method based on non-local mean filter

Abstract

A system and method for Bayer domain image noise reduction based on non-local means. First, the noise in the image is converted into Gaussian noise through variance-stabilized forward transformation; Pre-judgment, similarity weight calculation is performed for the reference points that pass the similarity prediction, and non-local mean filtering is performed on them, noise compensation is performed on the image that has passed the non-local mean filtering, and finally the variance is stabilized Transform to obtain a denoised image. Therefore, this method can remove the noise in the image and keep the texture details of the image well, so that the image looks more natural.

Description

technical field [0001] The present invention relates to an image processing technology, and further relates to a Bayer domain image noise reduction system and method based on non-local means, directly applied to the original RAW image collected by the camera, which can effectively eliminate the process of image acquisition and transmission noise generated in . Background technique [0002] Image noise is the part of the image that does not carry valid information. The destructive effect of noise on images is mainly manifested in the following two aspects: First, it affects subjective visual experience and reduces people's ability to watch, understand and appreciate images. For example, snow flakes on TV and blurry paper prints. The second is to make the follow-up middle and high-level research and application of the image impossible or to draw wrong conclusions. [0003] The noise in digital images mainly comes from the process of image acquisition and transmission. In t...

AI Technical Summary

Problems solved by technology

The second is to make the follow-up middle and high-level research and application of the image impossible or to draw wrong conclusions

Although the ability of these filters to maintain details has been enhanced, they still have common defects in image filtering: first, the acquisition of weights is based on a single pixel, and when these pixels are polluted by noise, the weight acquisition is ineffective. The stickiness becomes worse; the second is that there are often some false textures in the noise reduction results of these filters

Third, due to the use of image blocks to calculate pixel similarity weights, the computational complexity of the non-local mean algorithm is very large compared with the classical spatial domain algorithm

Fourth, the non-local mean algorithm is difficult to achieve a good balance between noise removal and image information texture preservation, especially for discrete strong noise in the image, or the noise reduction strength is too small to be effectively removed, or the noise reduction strength is too large to At the same time, a lot of effective image information is removed

Images