Research And Implementation Of Conference Video Super-resolution Technology Based On Deep Learning

The rapid development of computer,network and multimedia technology promotes the emergence of video conference.Video conference is an efficient and flexible conference method that is widely used in medical,military,business,and other fields,and conference video is a kind of video recorded by the camera equipment at the venue.Nowadays,video conference systems are becoming important communication services in various industries.At the same time,the video quality of conference videos has gradually attracted attention.Due to the limitation of network bandwidth or other recording hardware,it may not be able to record conference video with native high-resolution when recording conference video,and the visual effect is poor,which requires video super-resolution technology to improve the resolution of conference video.Video super-resolution technology refers to converting a segment of lower-resolution video into a higher-resolution video in some way.In recent years,video super-resolution technology based on deep learning has become a focused area in computer vision research.Deep learning-based video super-resolution technology builds a model by learning a large number of low-resolution and corresponding high-resolution video frames,and passes the low-resolution video frames through the learned model to restore high-frequency details of the frame.However,the application of video super-resolution technology to conference videos has not been satisfactory so far.Because when watching conference videos,the audience's focus is often on the human face,but the currently proposed video super-resolution technology can't reconstruct the face with high quality,and the overall reconstruction speed of the video is slow.In order to solve the above-mentioned difficult problems,this thesis has conducted in-depth research on conference video super-resolution technology.The main work includes the following:1)Aiming at the characteristics that the face is the focus of the viewer in the conference video,a full-convolution end-to-end deep neural network is constructed.The face fine module is introduced to reconstruct the face in the conference video.In addition,the optimization of the training process of the network is studied,and a training data enhancement method is proposed to make the network have better generalization ability.The comparative experiments have confirmed the advancement of the conference video super-resolution method proposed in this thesis.On the self-built conference video test set,the super-resolution frames generated by the network proposed in this thesis are better than other methods in objective evaluation indicators.,especially in the face reconstruction,the details are richer.2)Aiming at the characteristics of high similarity between frames and small motion scale in the conference video,a space-time complementary super-resolution technology is proposed,which replaces some super-resolution frames by using time super-resolution method.This method reuses the optical flow prediction network of the image registration module in the conference video superresolution network,avoids multiple training of the network,and uses the time super-resolution method with lower computational cost instead of the space super-resolution method with higher computational cost.Through experiments,it is confirmed that the method can achieve better results in most conference videos,and the overall reconstruction time of the video is greatly reduced without reducing the quality of the conference videos.