Research Of Self-supervised Real-world Image Denoising Via Multi-dimension Masking

Image denoising has always been one of the important research topics in low-level computer vision tasks.In the early days,there were some traditional denoising algorithms based on filtering and image priors.These algorithms can achieve well performance in Gaussian noise removal,but for real images taken by mobile phones or SLR,the performance is relatively weak.With the development of science and technology,especially artificial intelligence,more and more researchers have begun to study image denoising algorithms based on machine learning,and have proposed many excellent algorithms.These models need to use paired clean-noisy image data for training to achieve better results.However,it is extremely difficult to collect such paired clean-noise image data in real life.In order to solve the problem of paired data construction,researchers began to invest in the research of image denoising algorithms based on unsupervised learning and selfsupervised learning.This type of algorithm aims to use unmatched clean-noise image data or only use noisy images to train the denoising network to achieve better denoising ability.Among them,methods based on unsupervised learning usually use real noise to learn a noise generator and use clean noise to synthesize pairs of clean-noise image data pairs.The motivation of this method is very interesting,but it is often difficult for the noise generator to completely simulate the real noise distribution,so some new noise will be introduced,which will lead to deviations in model training and ultimately limit the overall performance of the denoising network.Self-supervised image denoising methods can be divided into two categories,the first is to construct training data based on the traditional self-supervised image denoising framework,and the other is to introduce some new framework structures such as noise decomposition network,using a prior of the noise distribution to construct a loss function to optimize the network.However,the existing image denoising algorithms based on selfsupervised learning have three serious problems,namely,limited noise removal ability,poor texture restoration ability,and moire artifacts in the denoised image.To address the above issues,we propose to employ cross-channel relationships to improve self-supervised image denoising.Recent self-supervised image denoising methods mainly utilize pixel-level data augmentation,such as pixel-level masks,pixelshuffle down-sampling,and submap samplers.With these augmentations,the model can exploit the spatial similarity of images,but cannot use cross-channel correlation,resulting in suboptimal performance.In order to solve this problem,we propose a new channellevel mask modeling strategy to enable the network to use cross-channel correlation.During the application,it was found that a single channel-level mask modeling would cause color distortion,so we further propose hybrid channel-level mask modeling and finally present multi-dimensional mask modeling.The multi-dimension mask proposed in this paper includes hybrid channel-level mask modeling and pixel-level mask modeling,which can comprehensively utilize spatial similarity and cross-channel correlation.During the training phase,we augment the original noisy image by randomly sampling two strategies from multi-dimension mask modeling.A large number of experiments show that the proposed method is superior to the existing image denoising methods based on selfsupervised learning,which proves the effectiveness of the proposed method.