Single Image And Video Super-Resolution Reconstruction Method Based On Convolutional Networks

Current super-resolution algorithms for single images and videos are widely used in areas such as medical imaging,video surveillance and security,remote sensing,etc.,and have been the focus of academic research because of their pivotal applications and value.In this paper,we make use of the deep learning techniques that have made great achievements in image and video processing over the years,and combine them with deep learning methods to achieve the task of super-resolution reconstruction of single images and videos,and the research mainly includes the following aspects.(1)In response to the current mainstream algorithm in which the network structure is getting deeper and deeper leading to the reconstruction accuracy and speed not achieving both the effect,this paper proposes a lightweight single-image super-resolution reconstruction structure combining residual network and target detection RFB.The network can extract richer and more detailed characteristic information of the image at a deeper level.It is shown that the cavity convolution in the RFB network is an effective way to capture the image to increase the perceptual field,while the residual network can effectively deal with the network gradient disappearance and explosion problem.In the image reconstruction module,pixel rearrangement of channels is performed by embedding sub-pixel convolution.The biggest advantage of sub-pixel convolution compared with transposed convolution is that the receptive field of neurons is large,which can provide richer contextual information to the super-resolution reconstruction module and obtain better reconstruction results.(2)In response to the fact that most video super-resolution methods rely heavily on the accuracy of motion estimation and compensation,and in order to solve problems such as artificial artifacts caused by heavy reliance on motion compensation,this paper proposes a new end-to-end deep neural network that can generate reconstruction frames by implicitly and dynamically upsampling filters and residual networks,and calculate each pixel according to its local spatio-temporal neighbourhood in order to avoid explicit motion compensation.Using our method,an HR image is reconstructed directly from the input image using the dynamic upsampling filter and added to the details computed through the RFB residual network structure in the image hyper-segmentation reconstruction,effectively exploiting the inter-frame temporal information for better feature fusion and achieving super-resolution reconstruction of the target frame.Compared to previous methods,the network in this paper is able to reconstruct sharper and more time-coherent high-resolution video with the aid of a new enhancement technique.We validate the effectiveness of the network through multiple sets of experiments to show that the network performs implicit processing of motion.In summary,for image super-resolution from both single image and video,the two models can be practically applied to the field of remote sensing medicine and have good practical applications.