Research Article
Deep Learning Based Video Compression
@INPROCEEDINGS{10.1007/978-3-030-99188-3_8, author={Kang Da Ji and Helmut Hlavacs}, title={Deep Learning Based Video Compression}, proceedings={Intelligent Technologies for Interactive Entertainment. 13th EAI International Conference, INTETAIN 2021, Virtual Event, December 3-4, 2021, Proceedings}, proceedings_a={INTETAIN}, year={2022}, month={3}, keywords={Deep learning Video compression Video reconstruction}, doi={10.1007/978-3-030-99188-3_8} }- Kang Da Ji
Helmut Hlavacs
Year: 2022
Deep Learning Based Video Compression
INTETAIN
Springer
DOI: 10.1007/978-3-030-99188-3_8
Abstract
Our goal is to test the capability of deep learning for compressing the size of video files, e.g., for sending them over digital networks. This is done by extracting keypoint and affine transformation tensors, using a pre-trained face model and then reducing the data by quantization and compression. This minimal information is sent through a network together with full source images used as starting frames for our approach.
The receiver device then reconstructs the video with a generator and a keypoint detector, by transforming and animating the keypoints of the source image according to the video keypoints. We minimized the required data by using LZMA2 compression and a quantization factor of 10 000 for keypoints and 1 000 for transformations.
Lastly, we determined limitations of this approach and found that in regard to file size reduction, our approach was noticeably better, while the quality of the resulting video in comparison to the original one was only half as good.
