Architecture of Deep Neural Network-based Video Compression

In our previous article, we introduced the status of Deep Neural Network-based Video Coding (DNNVC) approaches in the Moving Picture Expert Group (MPEG), one of the most important standardization groups for video compression technologies. In principle, video compression systems seek to minimize the end-to-end reconstruction distortion under a given bit rate budget, called a rate-distortion (R-D) optimization problem. To this end, a lot of efforts in video compression had been focused on the development of video coding tools, such as prediction, transform, entropy coding, and visual quality enhancement. These tools are devised to exploit spatial, temporal, and statistical redundancies in video signals.