Prior Based Image Restoration Research

In recent years,with the rapid development of electronic information technology,various imaging devices,such as digital cameras and smart phones,are popping up.These imaging devices not only enrich our daily life,but also play an important role in many fields.On one hand,due to the influence of different environmental factors(such as lighting changes and weather condition)and the physical limitations of the acquisition equipment,digital images will inevitably contain different types of degradation during the imaging process.On the other hand,the content of digital images is complex,thus behaving different statistical characteristics.Therefore,it is important that we can design efficient but adaptive image restoration algorithms to process the image content with different statistical characteristics according to different application requirements,and realize the adaptive enhancement of image/video in practical applications.Based on the rich statistical property and prior information of natural images,this thesis studies the tasks of image denoising,image super-resolution and image quality assessment,and focuses on the typical research problems in the field of image restoration.The main works and contributions of this thesis are summarized as follows:· Starting from the local prior of images,this thesis studies the classical bilateral filtering algorithm on image denoising task,and attempts to solve the problems of parameter selection sensitivity and poor robustness of bilateral filter.A new robust similarity distance metric is proposed due to the problem that the bilateral filter is not robust at high noise levels.For the problem that the bilateral filter is sensitive to parameter selection,based on the statistical characteristics of the images,we propose a content-aware adaptive parameter selection mechanism.· Based on the rich non-local similarity prior in the image,we proposed a non-local based dual domain denoising algorithm and design a general principal component analysis based denoising enhancement framework to improve the existing denoising methods.Firstly,transform based denoising methods usually generate ringing effects due to Gibbs phenomenon,which will reduce the visual effect of images.To address this issue,a Foveated non-local dual-domain denoising algorithm is proposed.The core of the method is to decompose the image into low-frequency and high-frequency parts.The low-frequency parts are first filtered by Foveated non-local means algorithm to make the filtered images smoother.The high-frequency parts use wavelet denoising to further remove the noise.Secondly,by combining the non-local similarity and low rank prior,a general denoising enhancement framework is proposed to enhance the existing image denoising algorithm.· We use deep learning to learn the prior information of the image from the training samples,and propose a second-order attention network for single image superresolution.Specifically,the proposed network mainly utilizes the spatial and channel correlations of the feature layer.· It is difficult to evaluate the image quality without reference images.For example,the images generated by image restoration algorithms usually contain multiple types of degradation(such as blur,noise and blockiness distortion).For this reason,a noreference image quality metric for multiply-distorted images based on structural degradation is proposed.The metric is based on the fact that the human visual system is sensitive to image structures.