Image denoising method based on combination of multiple scales and generalized kernel norms

Abstract

The invention relates to a multi-scale generalized kernel norm-based image denoising method, which comprises the following steps of: partitioning a natural image data set D, calculating mahalanobis distances corresponding to pixel distribution of image blocks for clustering, and determining an external prior parameter theta; internal image blocks are guided to be grouped, noise images Y under different scales are decomposed into N image blocks, and the noise images are divided into K similar block groups by calculating the mahalanobis distance; and constructing a regular term R (Z), performing low-rank constraint and noise filtering on each similar block group in the grouping stage, performing image integration on denoised image blocks under different scales, averaging overlapped parts, and obtaining a denoising result X with the same size as an input noise image. Noise image information between different scales is considered, statistical distribution rule characteristics of external clear images are also concerned, high robustness is achieved for complex image edges and textures, and noise and artifacts are restrained to a certain extent.

Description

technical field [0001] The invention relates to the technical field of image processing, in particular to an image denoising method of multi-scale generalized kernel norm. Background technique [0002] With the rapid development of Internet information technology, images are widely used in various industries due to their large amount of information and strong intuitiveness. In reality, digital images are affected by imaging equipment and external environments during digitization and transmission, which will lead to image quality degradation and limit subsequent visual tasks, such as target retrieval, target recognition, face recognition, video noise reduction and other applications. Remove Image noise is a very important preprocessing work in subsequent applications. Many scholars have deeply studied the application of image reconstruction tasks in the fields of computer vision and image processing. [0003] Image denoising methods can be divided into internal prior image d...

AI Technical Summary

Problems solved by technology

Patent CN110706181A discloses an image denoising method based on multi-scale dilated convolutional residual network. In the pre-training stage, the method builds a network model, adopts a combination of batch normalization and residual learning, and adopts the optimal Mix the expansion rate mode and introduce a multi-scale structure to obtain an end-to-end image denoising model. This type of method learns a certain mapping relationship between natural images and noise images based on deep learning, which requires a large number of training sets and has the ability to generalize poor

Note that the denoising method based on external priors makes full use of the prior information of natural images, but it is difficult to properly estimate the rank of similar block groups

Images