Super-Resolution Based On Progressive Iterative Back-Projection And Bidirectional Temporal-Recurrent Propagation

In the 21st century,the rapid development of multimedia technology has accelerated the popularity of HD monitors,and brought a higher and higher qualitive demand of image,which is the most important carrier of information,in the real-life applications.Considering the various degradations during acquisition and transmission of images,it’s urgent to figure out how to extract features of Low-Resolution distorted images and obtain the high-level visual language information to generate HighResolution images with plentiful texture details.Finally,it’s apparent to see the great value and significance in Super-Resolution technology.Efficient use of temporal and spatial features is the key to effective SuperResolution based on Convolutional Neural Networks,on which this article focuses.After analysis of problems in existing models,an improvement on spatial feature extraction and sampling mechanisms of Single-Image Super-Resolution is proposed.Furthermore,on the basis of this improvement,this article introduces a Video SuperResolution method,which can handle the extraction and fusion of temporal and spatial features.The work mainly consists of two parts as follows:(1)Aimed at the unreasonable spatial feature extraction and sampling mechanisms in large-scale Single-Image Super-Resolution models,a progressive iterative backprojection network is designed to achieve up-sampling mapping with high efficiency.Specifically,this network implements iterative back-projection units,which close to the human visual system,to introduce the error-feedback mechanism.Dense connection is added to enhance the information flow from shallow layers to deep layers.The network structure is modified to bring the role of dense connection into full play.Furthermore,this model uses sub-pixel convolutional layer and its inverse transform rather than deconvolutional and convolutional layers as the more efficient sampling mapping method.Experimental results prove the ability of recovering regular patterns in proposed models,and substantial performance enhancements in large-scale SingleImage Super-Resolution.(2)Aimed at the inefficient use of temporal features in existing Video SuperResolution models,a bidirectional temporal-recurrent network is designed for the enhancement in the use of temporal and spatial correlations.Specifically,the proposed network is implemented in recurrent architecture,which uses the propagation of temporal feature information among frames to avoid the explicit motion compensation in most deep-learning based Video Super-Resolution methods.The bidirectional propagation mechanism improves the efficiency of the use of information among frames.The proposed network also brings the progressive sampling method mentioned in part(1)into use to avoid one-step large-scale up-sampling mapping.Furthermore,Channel-Attention mechanism is added to achieve the channel-wise feature scaling,which brings more accurate feature fusion.Experimental results show that the proposed bidirectional temporal-recurrent network brings higher efficiency in use of temporal information among frames and superiority in Video Super-Resolution.