Research On Image Restoration Based On Adaptive Deep Feature Filtering

In the era of rapid information development,images contain scenario-based and intuitive information.The technology of image processing has also achieved practical applications in many fields,such as unmanned driving,intelligent security,medical diagnosis,etc.However,digital images inevitably generate noise during acquisition,compression,transmission,etc.,which leads to the degradation of the visual quality of the images.The appearance of such phenomena has a huge impact on subsequent computer vision processing.Therefore,the research on image restoration techniques is far-reaching.Such images not only affect the subjective visual understanding of humans,but also have a negative impact on subsequent computer vision algorithms when processing.Therefore,image restoration techniques have great practical application value.In this thesis,based on deep learning networks,the research is carried out for image denoising and single-image super-resolution reconstruction techniques,respectively,and the main research work of the thesis is as follows:1)The mainstream approach of real image denoising based on convolutional neural networks(CNN)consists of two sub-problems,namely noise estimation and non-blind denoising.In related work,the estimated noise prior is usually merged by channel merging and then a convolutional layer with a spatially shared kernel is used to achieve fusion of the noise prior with image features.Due to the variable noise intensity and feature details at all feature locations,this design cannot adaptively adjust the corresponding denoising patterns.To solve this problem,a novel conditional filter is proposed in this thesis,in which the optimal weights of different feature positions can be inferred adaptively from the local features and noise mapping of the image.Second,we bring the idea of alternating noise estimation and non-blind denoising into the network structure to achieve constantly updated noise prior to guide image denoising.In addition,a new affine transform block is designed to predict the stationary noise component and the signal-dependent noise component based on the properties of the heteroskedasticity Gaussian distribution.This thesis conducts a large number of experiments on mainstream datasets to verify the advantages of the denoising model in this thesis.2)Single-image super-resolution methods often use gradient prior information to guide the reconstruction.Among them,the gradient-guided fusion is achieved by channel merging and convolution layers.However,the convolutional kernel shared over spatial locations cannot be adapted to the gradient-guided effect at all feature locations.To this end,a new module is proposed to simulate the conventional joint trilateral filter by extending the definition domain from the pixel domain to the feature domain.Second,a new hyper-segmentation framework is built based on this module to infer high-resolution image features and gradient features simultaneously within two parallel branches,respectively.Among them,by using image features and gradient features as cross-guides to each other,the new module proposed in this thesis adaptively adjusts local features for fusion.In this thesis,through extensive experiments,the hyper-segmentation model in this thesis shows significant gains compared with other methods.