Research On Image Denoising Based On Transformer

In today’s intelligent society,digital images are widely used and are an important link for human interaction and human-computer interaction.However,in production and life,images will be affected by different factors in the process of recording,dissemination,and acquisition,which will introduce image noise,reduce image quality,and affect image content.Therefore,image denoising has important practical significance.Since the development of image denoising technology in the1960 s,it has been a hot issue in the computer field and has been widely used in medicine,aerospace,agriculture,night shooting and other fields.In the current image denoising technology,the deep learning denoising algorithm based on convolution occupies the mainstream position,but the denoising algorithm based on Transformer has achieved the best denoising effect.At present,the denoising algorithm based on Transformer mainly uses Swin Transformer as the core.Swin Transformer still has certain limitations in the feature extraction process of the attention mechanism,and the calculation time is too long compared with the deep learning method.Therefore,this paper is based on the Transformer construction algorithm.Compared with the current leading Transformer-based denoising algorithm,it improves the denoising effect and reduces the calculation time.The specific research contents are as follows:(1)A cross-scale feature fusion Transformer denoising network is proposed.Firstly,aiming at the limitation of global self-attention caused by the window self-attention mechanism in Swin Transformer,a multi-scale framework is proposed to fuse the multi-scale features extracted by the Transformer module across scales.Secondly,a separable convolution module is introduced to improve the local representation ability of Transformer.A large number of experiments have proved that compared with the traditional Swin Transformer,CSformer has achieved a maximum 0.03 d B improvement in the denoising index PSNR,and at the same time reduced the calculation time of a single image by 84.5%～97.4%.(2)A dynamic multi-scale Transformer denoising network is proposed.Firstly,aiming at the global self-attention limitation caused by window self-attention mechanism,this chapter introduces the idea of multi-scale by integrating window self-attention module branches of different window sizes.On the other hand,due to the large differences between images,the simple fusion of different scale features makes the network generalization weak.Therefore,this chapter introduces the channel attention module to learn the global channel features of images and dynamically assign weights to different scale attention modules.Experiments have proved that compared with the current optimal denoising algorithm Restorer,DWformer has achieved a maximum lead of 0.15 d B in high-level noise denoising index PSNR,which can effectively reduce image artifacts and texture loss.(3)A cross-scale feature fusion channel attention Transformer denoising network is proposed.The self-attention module is the core module in Transformer,which widely extracts global features in space,but uses shared weights between image channels,and does not effectively learn the relationship between feature map channels.In response to the above problems,this chapter proposes a cross-scale feature fusion network based on channel attention Transformer.On the one hand,a channel attention module is introduced to integrate the learned image channel features into the feature map for subsequent learning.On the other hand,the multi-scale idea is introduced using the cross-scale feature fusion network proposed in CSformer.Relevant experiments have proved that compared with the CSformer algorithm,CSCAformer has achieved a 0.01 d B～0.03 d B improvement in the PSNR value of the denoising index while the calculation time is close.