Title: A content-adaptive video compression method based on transformer

Authors: Heting Li; Dongsheng Jing; Ping He; Jiasheng Wu

Addresses: State Grid Suzhou Power Supply Company, State Grid Jiangsu Electric Power Co., Ltd., Suzhou Power Supply Company, Suzhou, Jiangsu, 215004, China ' State Grid Suzhou Power Supply Company, State Grid Jiangsu Electric Power Co., Ltd., Suzhou Power Supply Company, Suzhou, Jiangsu, 215004, China ' State Grid Suzhou Power Supply Company, State Grid Jiangsu Electric Power Co., Ltd., Suzhou Power Supply Company, Suzhou, Jiangsu, 215004, China ' State Grid Suzhou Power Supply Company, State Grid Jiangsu Electric Power Co., Ltd., Suzhou Power Supply Company, Suzhou, Jiangsu, 215004, China

Abstract: Convolutional neural network architectures have been the primary choice for deep learning-based video compression algorithms in recent years, but common convolutional neural networks can only exploit local correlations, while compression is faced with a wide variety of data types, making the compression performance and generalisation ability of the model challenging. To address the above challenges, a transformer-based content adaptive video compression method (TAVC) is proposed in this paper, which can effectively improve the generalisation ability of the model while achieving better compression effects. Specifically, we exploit the non-local correlation between features and propose a transformer-based compression network for motion information coding and residual coding to further improve the performance of compression coding. In addition, we design a content-adaptive algorithm to choose the best encoder parameters for various videos. Experiments show that TAVC outperforms current mainstream deep learning-based video compression coding algorithms on the HEVC and UVG datasets, saving an average of 14.944% of the bit rate.

Keywords: deep learning; video compression; content-adaptive; transformer.

DOI: 10.1504/IJCSE.2024.139770

International Journal of Computational Science and Engineering, 2024 Vol.27 No.4, pp.495 - 503

Received: 18 Sep 2023
Accepted: 10 Jan 2024

Published online: 05 Jul 2024 *

Full-text access for editors Full-text access for subscribers Purchase this article Comment on this article