Video Super Resolution Reconstruction Based On Deep Learning

With the rapid development of multimedia information,digital images and videos are widely used in various fields of life,such as meteorological remote sensing,military operations,public safety and so on.However,the resolution of images and videos cannot meet people's demands due to various unavoidable factors which include the imaging system,the manufacturing level of optical devices,the external factors and the impact of transmission equipment and so on.To get high-resolution videos and images,we improve the resolution of image and video mainly from the aspect of software,this process will use the image and video super-resolution method.Image super-resolution is an image processing technology,which recovers a high-resolution image from a low-resolution image.Video super-resolution is based on image superresolution,and it is equivalent to super-resolution of multi-frame images.The most important feature of video is the correlation between video frames.The existing video superresolution methods can be divided into two categories: motion estimation method and selfsimilarity method.The motion estimation method can only be applied to the situation that the motion of the object between the two frames is the translation movement and the displacement is not very large.The self-similarity method can only make a rough judgment on the self-similarity of the images,and cannot include all cases.The above methods have high computational complexity and poor performance.At present,many deep network models with different structures have made great progress in the reconstruction precision and computational performance of single frame and multi frame image super-resolution.The main work and contribution of this thesis are:1.This thesis proposes a space-dimensional video super-resolution algorithm based on deep learning.Combining the convolution neural network and video super-resolution by analyzing and mining the inter-frame correlation characteristics of video images.Extracting the correlation between the video frame through the three-dimensional filter and highresolution images are reconstructed by sub-pixel interpolation method.The experimental results validate that the proposed algorithm can get more stable reconstruction results and more detailed information compared with the existing methods.2.Furthermore,this thesis proposes a time-dimensional video super-resolution algorithm based on deep learning,that is,apply frame interpolation to the video.In our network,the time-dimensional super-resolution video is regarded as a set of feature graphs of the network outputs.Our inter-frame interpolating filters not only use the relevant information between frames,but also use the local structure information within the frame by training the neural network,which can get high-resolution video on time domain.The experimental results show that the proposed method can obtain better interpolating effect,and the time-dimension information of the reconstructed video is more abundant.