Referenced Frame Generation Network for Learned Video Compression

Nowadays, video data is becoming the majority of internet traffic, which brings new challenges for video compression algorithms. Under this circumstance, learned video compression (LVC) has emerged as its ability of end-to-end optimization. However, most of existing works only used the latest previous reconstructed frame as the referenced frame, ignoring rich correlated information in other reconstructed frames. To address this issue, we propose a referenced frame generation network (RFGN) for LVC. The RFGN combines information from multiple previous reconstructed frames to generate a highquality referenced frame for the motion compensation in LVC. Experiments indicates that the Deep Contextual Video Compression (DCVC) integrated with the proposed RFGN achieves an average of $\mathbf{4. 5 9 \%}$ bit rate saving comparing with the baseline.