Research On Image Restoration Through Deep Convolutional Neural Network

With the development of science,technology,economy as well as the quality of life,cameras have become indispensable items in people's daily life.They usually use it to take large quantities of images and record what has happened around them.However,capturing high-quality images requires expensive camera hardware and suitable photographing technology.Due to the defects of the hardware in ordinary consumer cameras and the influence of external environmental,it is generally difficult to obtain subtle and satisfied images.Some images also suffer from the problems of noise or blur,and others usually have low-resolution.With the development of computer vision theory,a variety of image processing algorithms have also been used to improve image quality.In this dissertation,from the perspective of improving the quality of real-world images,we establish deep neural network-based models to accomplish the tasks of denoising and super-resolution towards low-quality real-world images so as to improve the image quality and promote the application of image quality improvement algorithms based on deep learning in actual scenes.Its main contents are as follows:Firstly,this article introduces the background and significance of real-world image quality improvement,overviews various image denoising and super-resolution reconstruction methods,and proposes the research methods and key problems to be solved.Secondly,this thesis introduces the imaging process in a camera of real-world images and image quality assessments,outlines the mathematical model of image degradation,and lays a theoretical foundation for image reconstruction.Thirdly,based on the real-world image denoising task,starting from the imaging principle of the biological vision system,an algorithm aiming at adjusting the scales of receptive field dynamically is proposed.According to the targets in different regions of the image,the mechanism of adaptive selection of multi-scale receptive field is realized through the Softmax function.At the same time,we propose the de-pixel-shuffle operation,which aims at reducing the spatial size of feature maps at the cost of adding multiple channels so as to reduce the feature redundancy of the image,accelerate the execution speed as well as improve the denoising results.Fourthly,as for the super-resolution task,starting from the difficulty of image reconstruction of real-world images,we propose a novel single image super-resolution algorithm via channel attention-based fusion of orientation-aware features.In this method,present a novel feature extraction module containing a number of well-designed 1D and 2D convolutional kernels to extract orientation-aware features.Then we design channel attention-based fusion schemes,which can adaptively combine features extracted in different directions and in hierarchically stacked convolutional stages.Finally,in order to prove the effectiveness of our proposed models,we implement some classical or novel denoising/super-resolution methods training on real-world images,and compare our image restoration results with theirs.According to the final experiments results,our proposed algorithms obtain the best PSNR/SSIM values and acquire the most subtle visualization performances among these state-of-the-art methods.