Research On Image Restoration Based On Deep Neural Network

With the emergence and development of deep neural network,the research process in the field of image restoration has been greatly promoted,which gave birth to many image restoration algorithms based on deep learning,and soon achieved far better performance than traditional methods.Among them,the research on image denoising task and image super-resolution task is the most popular.However,there are still some problems in the existing research of the two.For example,the former is generally based on Gaussian white noise training and has poor generalization performance in dealing with complex and changeable real noise,while the latter’s advanced model consumes more resources and has high calculation cost,which is not conducive to the application of research results in industry.Therefore,using deep learning technology,this paper focuses on the real denoising task and lightweight super-resolution task,which have more practical applicability,and puts forward effective algorithms respectively.Compared with the existing algorithms,they are improved in objective indicators and subjective effects.The main contributions of this paper include the following three aspects:(1)For the image denoising task,a real denoising algorithm based on multi-scale Transformer is proposed,which can effectively remove the real noise and restore more image details.The optimization of the model structure design mainly includes the following three parts:firstly,the locally-enhanced window Transformer is used as the feature extraction module,which can not only take advantage of the Transformer architecture in capturing long-distance dependence,but also introduce more local context information.Secondly,the multi-scale residual module is introduced to construct parallel multi-scale flows,so that the model can not only obtain spatial information from high-resolution features,but also capture rich semantics from low-resolution features.At the same time,the original resolution scale is ignored,which significantly reduces the computational complexity of the model.Finally,the multi-scale feature fusion layer is introduced into the multi-scale residual module to strengthen the information interaction between scales,so as to improve the richness of acquired features.The experimental results on real noise data sets SIDD and DND show that the proposed algorithm achieves the same subjective and objective quality as the comparison algorithm,but the amount of calculation of the model is reduced by more than 90%.(2)For the image super-resolution task,we propose a lightweight superresolution algorithm based on multi-scale Transformer to further explore the generalization of multi-scale Transformer architecture in the image restoration domain.This algorithm model is optimized on the basis of our denoising model.The main improvements are as follows:first,the multi-scale residual module with upper and lower shunts is adopted to significantly reduce the middle feature dimension of the model.Thanks to the effectiveness of the multi-scale architecture,the model can significantly reduce the amount of parameters while maintaining performance.Furthermore,we adjust the feature fusion mode of multi-scale feature fusion layer and introduce a selective kernel module which can adaptively combine scale features,which not only reduces the amount of parameters,but also improves the performance.The experimental part indicates that the objective metrics and subjective effect of this algorithm on multiple benchmark data sets are better than the comparison algorithms,and also shows the great potential of multi-scale Transformer architecture in the field of image restoration.(3)In order to verify and evaluate the practical effects of the two algorithmic models proposed in this paper,an image restoration system based on B/S architecture is designed and implemented.The system provides three functional modules,including real denoising,super-resolution and real denoising&super-resolution,to meet different usage requirements.The system client adopts the cutting-edge development framework.Users can easily upload degraded images through the browser,and then select the desired processing mode to obtain the results,and download the restored images.The system server is mainly responsible for parsing various requests from the client,so as to execute the corresponding pre-processing logic for the uploaded image and invoke the corresponding deep learning model.The overall design of the system is lightweight,with good user experience,low coupling between modules,easy to expand the system functions,and has certain practical value.