Image Super-resolution Research Based On Deep Learning

Image super-resolution technology refers to increase the spatial resolution of image and restore higher-frequency details to obtain richer image content by the means of software algorithms.Nowadays,image super-resolution technology has gained more attention in the field of video processing,and video super-resolution has become a research hotspot.Thanks to the advantages of low hardware cost and small deployment difficulty,video super-resolution technology can be embedded in various stages such as recording,transmission,playback,and application,where the technology plays a prominent role and have extremely high application value.We focus on the research of video super-resolution algorithm.Recent years,deep learning technology has received extensive attention and development on the problem of video super-resolution.According to the different ways of processing temporal information,it can be divided into multi-frame network and recurrent network.The multi-frame network obtains temporal information from adjacent multi-frame input images according to the autoregressive model,so it is limited by the number of input image frames and it is difficult to maintain inter-frame consistency due to an independent calculation process.The recurrent network obtains temporal information through the loop processing step by step,so it is limited by the number of loop processing,which leads to poor super-resolution quality in the initial stage.Regarding the problems above,we reform the multi-frame network and the recurrent network respectively,then integrates the multi-frame network and recurrent network to obtain a highly robust video superresolution method that maintains better inter-frame consistency.The main work includes:(1)For multi-frame network model,the image registration module is ameliorated by deformable convolution combined with non-local networks,which makes better accuracy of image registration;and the feature fusion module is ameliorated by the spatiotemporal attention mechanism of the progressive fusion structure,which strengthens the spatiotemporal feature fusion effect.By integrating the above modules,we design a video super-resolution network model based on non-local deformable convolution and achieve better super-resolution reconstruction quality.(2)For the recurrent network model,the information flow circulation method is ameliorated by the image processing result in the pixel domain combined with the spatiotemporal feature map in the feature domain to perform information circulation,which reinforces the ability to express time domain features and reduces the impact of error propagation;the network is ameliorated by integrating structures such as image registration module and feature fusion module,which reinforces the ability to express spatial features and the comprehensive ability of spatiotemporal features.We design a video super-resolution network model based on frame and feature information recurrent,which reduces the impact of error propagation and achieves better super-resolution reconstruction quality.(3)We design a robust video super-resolution method by integrating the above-mentioned ameliorated multi-frame network model and recurrent network model.By integrating the multi-frame network and the recurrent network,the overall expression of the spatiotemporal features of the network is further intensified,which helps to restore more high-frequency details;at the same time,it fully utilizes the advantages of both sub-network,which enhances the inter-frame consistency and gets better quality of the super-resolution image at the initial stage.The comparative analysis of the experimental results proves the effectiveness of the video superresolution methods designed in this paper,which has achieved certain improvements in both subjective visual effects and objective evaluation quality;the comparative analysis of the network structure ablation experiment further proves the effectiveness of enhanced network module and the integration methods in this paper.