An Efficient Video Compression Network
Prasanga Dhungel, Prashant Tandan, Sandesh Bhusal, Sobit Neupane, Subama Shakya · 2020 2nd International Conference on Advances in Computing, Communication Control and Networking (ICACCCN) · 2020
The proliferation of deep learning has catalysed the renaissance of video compression, as many frameworks yielding comparable or even better performance than conventional video codecs have been proposed in the past few years. Despite the improvement in rate-distortion, these models are much slower and require more memory which limits their practical usage. So, in this paper, we present a computationally efficient, deep learning based video compression framework. Specifically, we refine the shortcomings of conventional approach of video compression by substituting each traditional component with their minimalistic neural network counterpart. We adopt rate-distortion optimization as a primary objective and jointly optimize all the components under a scrutiny of a single loss function where the network learns to exploit the spatio-temporal redundancy present in the frames and reduces the bit rate while preserving visual quality in the decoded frames. Experimental results show that our approach generalizes well when trained on generic video content and achieves comparable performance than the widely used commercial codec AVC/H.264, and slight drop in performance compared to HEVC/H.265 and current deep learning based frameworks but in the expense of significant improvement in speed and memory requirements. We have released our code and pre-trained models at https://github.com/tukilabs/Video-Compression-Net for further research and improvement.