Research On Image Gaussian Denoising Algorithms Based On Deep Networks

Image denoising is a long-standing crucial problem in computer vision research,as the outcome of the subsequent image information processing is considerably limitted by the quality of the basic image.Boasting specific features,Gaussian noise is regarded as the most common one among the miscellaneous types of noises,so we gather our attention to the Gaussion denoising problem in this paper.Traditionally,the thinking of image prior modeling has long been considered as the pillar in image Gaussian denoising.This approach excessively relys on the experience of users and may neglect some high-level features during the modeling process.Recently,with the prosperity of deep learning and the advancement of hardware performance,image Gaussian denoising methods based on deep neural networks have demonstrated their excellent performance and have gradually become the mainstream in denoising tasks.Nevertheless,there are still many deficiencies with image Gaussian denoising methods based on deep networks.First,the current prevailing image Gaussian denoising methods based on deep neural networks do not take full advantage of the latest fruits from deep learning that have fittingly applied in pattern recognition,so there is still sufficient room for better performance.In addition,their application range is still greatly restricted by the specialized task(i.e.,a specific model is required for each considered noise level),which means the current deep denoising network lacks exceptional blind image Gaussian denoising capabilities.Finally,the quantitative evaluation index for the denoised image may be inadequate.Usually a slight increasce in the quantitative evaluation index value of the denoised image cannot be reflected in the improvement of its human eye quality.The current insufficient research on image Gaussian denoising methods prompts us to make amelioration of deep denoising networks in two directions.On one hand,we seek to enhance the structure of image Gaussian denoising networks by the absorption of the latest achievements in deep learning.Consequently,we propose a deep image Gaussian denoising network based on residual learning.Extensive experimental results have corroborated its superiory of image Gaussian denoising quality with pontential to deal with blind image Gaussian denoising tasks.On the other hand,we no longer blindly pursue the quantitative evaluation index of image Gaussian denoising,and aim to improve the quality of visual perception of denoised images.To this end,a deep denoising network is designed based on generative adversarial networks and perceptual loss.Empirical verification shows that our proposed method has the merits to both achieve the denoising quality equivalent to the mainstream image Gaussian denoising method and better retain the texture details of the denoised image.