| As one of the main information carriers in the era of multimedia,video has become the mainstream of Internet data in the time when big data is the trend.With the COVID-19 haunting the world,people's demand for real-time communication using online video for life,education,and work has become a must,and the demand for better video experiences is also increasing.Considering that it is impossible to transmit uncompressed video in real scenarios,researchers have always regarded it as the ultimate goal of the development of video coding standards to preserve the video image quality as much as possible while achieving the largest data compression ratio.At present,when high-efficiency video coding standards are constantly evolving in order to adapt to network status and user terminals,the deep learning technique led by convolutional neural networks has also made breakthroughs in the field of computer vision.Under the aforementioned background,this article starts with the basic principles of video coding,and analyzes the reasons that lead to the compression damage of video images.We focus on the unique multi-level and multi-correlation characteristics of spatial scalable video coding.According to the coding reference prediction method of traditional coding,we take the idea of multi-adjacent frames and multi-level collaborative reference as,and adopted the deep convolutional neural network technique,proposing a video spatial scalable coding algorithm based on coding damage repair and image super-resolution.Our proposed algorithm is based on the past research on image super-resolution and image coding damage repair,and is an improved scheme for spatial scalable coding.Based on the characteristics of the mapping relationship between the BL and EL layers and in between the individual levels,utilizing motion compensation and image super-resolution restoration,a motion compensation module and a resolution restoration module are proposed separately to realize a multi-neighboring frames and multi-level reference collaboratively.On the one hand,motion compensation makes full use of the highly correlated logic between the images that are of the same level sequence of frames,on the other hand,the latter fully mines the correspondence between the missing information before and after the image is compressed at the resolution level,and eventually the information extracted passes through a fusion mechanism to be combined,realizing the extraction of useful information that takes both aspects into account.At the same time,basing on previous studies,this paper reselected some high-quality video sequences,designed and produced a data set that meets the actual spatial scalable coding scenario,which complements the deep neural network proposed in this paper to aid in the network training,achieving a better result.Our experimental results show that,compared with the reconstructed image with compression coding impairment,the PSNR gain of the inner layer frames restored by the training network in this paper can reach up to 2.361 d B,and the average value of the BD-BR reduction ratio can reach 6.007%.From the perspective of subjective evaluations,the image detail information that has been distorted/disappeared in the reconstructed images has been restored to a certain extent,and the naked eye effect is better.The inner layer frames are the video frames referenced by the enhanced layer in spatial scalable coding structure.The improvement of their quality means a better reference for the enhanced layer code procedure,which not only improves the coding effect of the enhanced layer,but also improves the visual quality of the reconstructed video.The overall image quality of the code stream is also reduced to a certain extent,which improved coding efficiency,and eventually we achieved the ultimate goal of this experiment to reduce the burden of network transmission and adapt to heterogeneous network conditions and terminal equipments. |
PDF Full Text Request