Real Image Denoising Based On Unsupervised Learning

With the development of digital multimedia technologies,people's requirements for image quality are increasing day by day.However,real images are always inevitably polluted by noise during the process of acquisition,compression,transmission,and display.Therefore,how to reduce noise pollution by using algorithms from existing real noise images while retaining valuable feature information of the image as much as possible has very important theoretical research and practical application value.Although the current image denoising algorithm for synthetic noise has achieved very good results.However,because the types and distribution of noise contained in real camera images are more complicated,and it is difficult to collect large-scale noise-free images as training samples,it is difficult to apply image denoising algorithms based on synthetic noise directly to real image noise removal.To reduce the model's dependence on data samples and give full play to the inference ability of deep learning,this paper takes unsupervised learning as a starting point and conducts in-depth research around the application of real image denoising.In this paper,according to the image degradation process of additive noise,we propose an unsupervised learning framework DIPDN(Double Image Prior Denoised Network)for blind denoising of real images.We use two convolutional generation networks to learn the image layer information and noise layer information of noise image respectively.In this framework,we use the reconstruction loss to guide the network to learn the degradation process of a single noisy image and additionally construct an image prior constraint and a noise prior constraint to guide the training of the two subnetworks,respectively.Besides,for the problem that the generation network is easy to cause the loss of image detail information,we propose a progressive backprojection generation network to retain and restore more image detail textures.Through experimental verification on the synthetic noise data set and the real noise data set,the DIPDN proposed in this paper achieves a higher value of evaluation index objectively,and the denoised image is more consistent with the original image.Subjectively.According to the common continuous shooting mode on mobile phones or other camera equipment,we also propose a multi-frame image fusion denoising model MIPDN(Multi-frame Image Prior Denoised Network)based on unsupervised learning based on DIPDN,which makes full use of the layer correlation and noise randomness between multi-frame images for fusion denoising,which significantly improves the image denoising effect.To solve the problem of DIPDN,MIPDN proposed three key improvements: noise network based on residual learning,filter response normalization unit,and training data enhancement scheme.A large number of experimental results show that the MIPDN denoising model achieves a high objective evaluation index and a good application effect in practical applications.Through the research in this paper,we propose a blind denoising algorithm for real images based on unsupervised learning,which combines the prior information of the images and the reasoning ability of deep learning and effectively improves the image quality of real noise images.