Exploring Invertible Encoding for Deep Video Compression

Abstract

Deep video compression methods typically use autoencoder-style networks for encoding and decoding, which can result in the loss of information during encoding that cannot be retrieved during decoding. To address this issue, recent work has explored the use of invertible neural networks for enhanced invertible encoding, which has successfully mitigated spatial information loss for better image compression. In this paper, we propose a new approach that extends invertible encoding to temporal information and introduces an encoding-decoding network for deep video compression. Our network incorporates a novel attentive channel squeeze module to improve compression performance while also leveraging a conditional coding framework for motion compression. The entire framework is optimized via a single loss function that balances bit cost and frame quality. The experimental results demonstrate the effectiveness of our approach, which achieves 15.45%/57.92% bit savings in terms of PSNR/MS-SSIM compared with the high-efficiency video coding low-delay P configuration.