HDVC: Deep Video Compression With Hyperprior-Based Entropy Coding

In this paper, we propose deep video compression with hyperprior-based entropy coding, named HDVC. The proposed method is based on the deep video compression (DVC) framework that replaces traditional block-based video compression with end-to-end video compression based on deep learning, aiming to improve compression efficiency and reduce computational complexity while maintaining visual quality. Based on the DVC framework, we introduce hyperprior-based entropy coding into motion compression and optimize motion vector estimation (i.e. optical flow estimation) using window attention and fast residual channel attention. Moreover, we introduce residual channel attention intermediate module into both encoding and decoding to enhance residuals and the quality of reconstructed frames. We adopt hyperprior-based entropy coding in residual compression to model feature distribution. Besides, we use learned image compression for intraframe coding based on fast residual channel attention network to generate reference frames. Experimental results show that the proposed method achieves better PSNR and MS-SSIM performance than both traditional block-based and recent deep learning-based video compression methods on UVG dataset.