Research On Low Quality Image Enhancement And Restoration Algorithms Based On Deep Learning

As an important multimedia information medium,image plays an important role in human visual perception.High quality image carries plenty of information,which can not only give viewers a pleasant visual experience,but also help to improve the performance of computer vision applications.However,due to the limitation of hardware equipment and the influence of imaging environment,the acquired images often have different degrees of degradation,resulting in a variety of low-quality images,such as low-resolution images caused by the sensor resolution limitation of image acquisition equipment,low contrast images caused by environmental factors such as haze,low light and so on.In recent years,deep learning technology has made continuous progress in image classification,target recognition and other computer vision fields.In the problem of low quality image enhancement and restoration,deep learning based methods are also emerging.According to the types of driving data,these methods can be divided into: reference data-based method,synthetic data-based method and no reference data method.Based on the deep learning theory,this thesis explores the enhancement and restoration of low-quality images,and proposes a variety of methods.The main work includes the following aspects:(1)On the issue of low-quality image enhancement and restoration based on reference data,this thesis focuses on the task of depth map super-resolution,and proposes the hierarchical features driven depth map super-resolution.On the basis of encode-decode network,the algorithm extracts multi-scale depth map features through input pyramid,integrates auxiliary information of guiding image layer by layer,makes full use of global and local features,realizes super-resolution of depth map and produces clear and sharp edges in the results.(2)In the issue of low-quality image enhancement and restoration based on synthetic data,this thesis focuses on the task of image dehazing,and proposes the perception-inspired single image dehazing network with refinement.The proposed dehazing network is a two-stage network,which consists of a haze removal sub-network based on perceptual loss and a refinement sub-network based on combined loss.The algorithm further enhances the dehazed result by the refinement sub-network to obtain the final result which is clear and more natural,and adapt to all kinds of real scene haze images.(3)In the issue of low-quality image enhancement and restoration without reference data,this thesis focuses on the task of low light image enhancement and proposes a low light image enhancement method based on the non reference depth curve estimation.Inspired by the image editing software,the proposed algorithm transforms the image enhancement problem into the mapping curve estimation problem,and drives the network learning curve parameters through a set of non-reference loss functions.It reduces the burden of the network and relaxes the requirements of training data.The algorithm can not only significantly enhance all kinds of low light images,but also greatly improve the performance of computer vision applications(such as face detection).