Learning to compress videos without computing motion

With the development of higher resolution contents and displays, its significant volume poses challenges to goals acquiring, transmitting, compressing, displaying high-quality video content. In this paper, we propose a new deep learning compression architecture that does not require motion estimation, which is most expensive element modern hybrid codecs like H.264 HEVC. Our framework exploits regularities inherent motion, capture by using displaced frame differences as representations train neural network. addition, space-time reconstruction network based on both an LSTM model UNet model, call LSTM-UNet. The has three components: Displacement Calculation Unit (DCU), Compression Network (DCN), Frame Reconstruction (FRN). DCU removes need for estimation found in less expensive. DCN, RNN-based utilized compress well retain temporal information between frames. LSTM-UNet used FRN learn differential videos. experimental results show our MOtionless VIdeo Codec (MOVI-Codec), learns how efficiently videos without computing motion. experiments MOVI-Codec outperforms Low-Delay P veryfast setting coding standard exceeds performance global HEVC codec, same setting, measured MS-SSIM, especially latest H.266 (VVC) codec at bitrates, when assessed high-resolution