Research On Image Restoration Method Based On Deep Learning

Image has become an indispensable part of our life.However,the image is easily affected by the surrounding environment and human factors during the imaging process,which leads to the degradation of image quality and affects the performance of some highlevel visual tasks.To improve the quality of images and provide clear images for subsequent high-level tasks,image restoration task has been proposed.In recent years,the image restoration task has been a hot research issue in the field of computer vision.With the development of deep learning,image restoration algorithms based on deep learning have been proposed and achieved excellent performance.However,these algorithms still have some problems.For example,the details of the reconstructed image are lost and the structure of the reconstructed image is not clear.Therefore,this paper focuses on how to protect image details and we will protect the details in three parts:(1)Introducing a model that can sufficiently utilize the intrinsic properties of features?(2)Making full use of prior information to protect the reconstructed image details?(3)High-frequency information represents the structure information of the image,while low-frequency information represents the content information of the image.Therefore,focusing on learning the different frequency information of the image to protect the details.In addition,we will demonstrate the effectiveness of the above three parts in three tasks(image dehazing,image deraining,and MRI reconstruction).The details are as follows:(1)An efficient image dehazing algorithm based on a multi-scale topological network is proposed,which mainly focuses on the extraction and utilization of image features.Specifically,the algorithm first adopts a topological network structure as the backbone.This topological network structure provides a large number of propagation paths to promote the transmission and utilization of features,so that the network can extract rich features and make the network robust and fault-tolerant.In addition,there is a strong correlation between different scale features.In order to make full use of features of different scales,the algorithm proposes a multi-scale feature fusion module and an adaptive feature selection module,which can adaptively select important features and ignore interference information.Compared with the existing dehazing algorithm,this algorithm achieves better dehazing performance.(2)A structure-preserving rain removal algorithm based on residual channel prior guidance is proposed,which mainly focuses on the use of prior information.In particular,the algorithm first designs a wavelet-based multi-level module to effectively learn the background information in the rainy image.Meanwhile,the algorithm also proposes the residual channel prior feature extraction module and the interactive fusion module to extract the residual channel prior feature and use the residual channel prior feature to effectively guide the algorithm to reconstruct a clear rain-free image.This is the key step of the algorithm.It can highlight the structure information of objects in the rainy image and promote to generate a high-quality rain-free images.Moreover,the algorithm also adopts an iterative guidance strategy,that is,it gradually extracts clearer residual channel prior information from the intermediate results and adopt the clearer residual channel prior to guide the algorithm to remove rain again.Experiments on synthetic rainy dataset and real-world rainy images show that the algorithm can restore clearer rain-free images.(3)A MRI reconstruction algorithm of multi-scale Fourier Transformer based on frequency learning is proposed,which mainly focuses on the learning of different frequency information of MR images.This method focuses on repairing the low-frequency and highfrequency information of MR images under the supervision of low-frequency loss function and high-frequency loss function.Specifically,the algorithm designs a high-frequency learning branch and a low-frequency learning branch.Meanwhile,the algorithm also proposes a multi-scale Fourier Transform module(MFT)for learning non-local similarity.Furthermore,the algorithm also introduces multi-scale learning and cross-scale linear fusion strategies in MFT to sufficiently interact the information between the features of different scales similar image patches and strengthen the representation of features.Based on MFT,this algorithm also designs a residual multi-scale Fourier transform module to further strengthen the representation of different frequency features.The experimental results under different acceleration rates and different sampling pattern on different datasets show that this algorithm is superior to other advanced reconstruction algorithms.