Application Of Convolutional Neural Network In Image Restoration

As a kind of information carrier,images have been widely used in our society.However,images will inevitably loss quality during the generation and transmission process,which will introduce irreversible information loss,and will have adverse effects on other high-level visual tasks.The purpose of image restoration is the process of modeling these degraded processes to restore the degraded images.In the field of image restoration,image super-resolution and JPEG compression artifacts reduction have also attracted much attention.With the development of deep learning in recent years,using convolutional neural networks to handle these tasks is also a hot topic.Because the end-to-end training strategy can better deal with the potential relationship between the degraded image and the original image,it has achieved better results than traditional methods.Therefore,the work of this thesis is also based on the convolutional neural network.Specifically,the main work of this thesis includes:1.We first analyze the recursive module commonly used in current image superresolution research,and discuss their shortcomings,then we propose a new recursive module called memory recursive module(MRM).Our proposed module can memorize the output features of each stage during the recursive processing,and provide more intermediate information for the next recursive stage,which can perform reconstruction more effectively.At the same time,in the feature upsampling phase,in order to process the output features in each stage of MRM more effectively,we propose a new feature fusion method called Shuffle Conv.This module only uses the most relevant channels between the output features of each MRM to rebuild the feature maps,which can improve the performance of the network while greatly reducing the number of parameters.2.Most current deep learning-based JPEG compression artifacts reduction research is limited to a specific JPEG quality factor,which cannot get good reconstruction results when dealing with pictures of unknown JPEG quality factor.To solve this problem,we propose a new JPEG compression artifacts reduction method for multiple quality factors.In our proposed network,the quality factor of the JPEG compressed image is evaluated by a JPEG quality factor evaluation sub-network,then the obtained evaluation information and the original input will be used in a JPEG compression artifacts reduction sub-network.In this approach,our proposed method can reconstruct images with different compression levels using the same model,and can also get high reconstruction quality when dealing with unknown JPEG quality factor.3.Since most current deep learning-based image super-resolution algorithms only process uncompressed images,applying these methods to compressed images will greatly amplify the noise caused by compression.But in reality,most of the images are compressed by the JPEG algorithm,so those methods cannot achieve good results in reality.To handle this problem,we propose an image super-resolution method for JPEG compressed pictures based on our previous research work.Specifically,a JPEG compression artifacts reduction network is used to reconstruct the JPEG compressed pictures,and then an image super-resolution network would use that reconstruction information and original input to rebuild the high-resolution image.By using this strategy our network can achieve better reconstruction results than the general image super-resolution methods.