| In the Internet era,the content of online video is increasing day by day,and people have higher and higher requirements for its quality.Due to the large amount of raw video content data,the existing hardware storage devices and network transmission devices are not enough to undertake the important task of network video storage and transmission.Therefore,video coding technology is proposed to compress video data,so as to reduce the storage capacity of network video and the bandwidth occupied by its transmission.Video coding technology mainly compresses video from two aspects: data redundancy and visual redundancy.In order to better optimize the coding efficiency,video coding technology uses lossy compression technology to reduce the bit rate in exchange for a certain objective distortion.Therefore,the compressed video using video coding technology has a certain quality loss,and may even affect the viewing performance.In order to enhance the quality of compressed video,there are two mainstream methods,one based on traditional method and the other based on deep learning method.Based on the traditional method,the manual model is used to filter the compressed video,so as to improve its quality.However,the manual model based on traditional methods has some disadvantages,such as artificial design model,limited improvement of quality and so on.The method based on deep learning is to learn a model through big data.Compared with traditional methods,this method does not need manual design,can learn adaptively from a large amount of data,and can greatly improve the video quality.The quality improvement of compressed video belongs to ill posed problem,that is,the compressed video information with information loss is restored to the original lossless video information.The current methods based on deep learning are divided into two kinds according to the input.One is single frame enhancement algorithm,that is,single frame input and single frame output,and the other is multi frame enhancement algorithm,that is,multi frame input,single frame output or multi frame input and multi frame output.The single frame enhancement algorithm only depends on the spatial information of the input image to enhance the quality,while the multi frame enhancement algorithm further uses the time domain information.Thesis designs two model algorithms for single frame enhancement algorithm and multi frame enhancement algorithm to improve the quality of compressed video.The specific work is as follows:1.Single frame enhancement algorithm: Aiming at the problem of how to more effectively use a priori information to guide the recovery process of compressed video,an explicit self-attention-based multimodality network is proposed.Two kinds of prior information,the division structure of coding unit and the edge structure of image object,are used to guide the restoration process of compressed image.Considering that the prior information and the compressed image belong to different modal information,the attention mechanism is used to effectively fuse the two-modal information.The experiment is based on VTM7.0 with All intra configuration,and BD-Rate is reduced by7.24%.2.Multi frame enhancement algorithm: Aiming at the temporal alignment problem in multi frame enhancement algorithm.A multi-scale gradient back propagation optimization framework based on deformable convolution is proposed to stabilize model training and enhance video quality.The framework uses a multi-scale alignment operation to optimize the deformable convolution training instability problem from the gradient back-propagation mechanism.The experiment is based on HM16.9 with LDP configuration,and BD-Rate decreased by 25.6%. |
PDF Full Text Request