Image Denoising Method Based On Residual Learning

Image is one of the main ways for human acquire and exchange information.With the popularity of digital image devices,the range of applications of images is also expanding.However,due to the influence of external environment,acquisition equipment and storage media,images are inevitably affected by noise during the process of image acquisition and transmission,which reduces the quality of the image.Damaged images are detrimental to the transmission of information and directly affect subsequent higherlevel images processing tasks.Therefore,image denoising techniques have attracted extensive attention from researches.Image denoising is important as a pre-processing method for the subsequent application of images,and it has practical applications in numerous fields such as picture monitoring,old photo restoration and so on.Many image denoising algorithms have been developed and improved in recent years,but there are still some limitations.Using a single model to deal with different levels or types of noise is not effective.Both noise and detail are high-frequency signals,and it is easy to lose detailed features while removing noise.Most deep learning methods rely on supervised training of pairs of data,yet it is difficult to obtain the corresponding clean images from real noisy images.Noisy images are generated from clean images with noise interference and the difference between them(i.e.noise)is regarded as the residual.The research work of this paper is based on residual learning,which is summarized as follows:(1)An image denoising algorithm based on multi-scale adaptive residual learning is proposed.The algorithm uses different dilated convolutions to form contextual aggregation block that expand the receptive field and capture rich noise information from different scales.Deformable convolution is introduced to adjust the sampling position according to different noise to enhance the adaptive capability to different features.Spatial attention is used to enhance important features in large amounts of information.The algorithm is able to deal with the same type and different levels of noise,even real noise,and effectively improve the effect of blind denoising.(2)A dual-stream image denoising algorithm based on feature fusion is proposed.The model consists of the parallel structure of residual estimation sub-network,a detail recovery sub-network and a feature fusion module.The residual estimation sub-network learns the difference between noisy image and clean image,which is able to remove noise from the input image to achieve initial denoising.The detail recovery sub-network adapts to changes in details such as edges and textures in the image,preserving the detailed features of the image.For the output of the two sub-networks,a feature fusion module is designed to refine the multi-scale semantic information and combine the effective features of each.The model provides a better balance between noise removal and detail recovery than existing methods.(3)A self-supervised image denoising algorithm based on generative adversarial networks is proposed.The model consists of two parts,the generator and the discriminator.The generator is based on residual learning and uses an up-down self-guided strategy.The dual self-attentive module is introduced to aggregate local features and global dependencies,which focuses on guiding the network to learn the noise information in the input image.On the basis of unpaired images,the use of discriminator and self-consistent loss function for phased adversarial learning is beneficial to reduce the interference of image features in the residual and improve the accuracy of noise estimation.The algorithm achieves image denoising under self-supervision,which can use the noise mapping generated by the generator to construct paired data and expand training samples.