CD-DVC: Conditional Diffusion Model for Distributed Video Coding

Conventional predictive coding-based video compression techniques encounter notable difficulties when deployed on wireless terminal devices with constrained computational capabilities, primarily due to their dependence on resource-intensive encoders. Distributed Video Coding (DVC) alleviates this issue by transferring the computational load from the encoder to the decoder. However, traditional DVC methods still fall short in performance when compared to predictive coding strategies. Drawing inspiration from the impressive advancements of deep learning in image and video compression, we introduce CD-DVC, an innovative end-to-end DVC framework. CD-DVC adheres to the independent encoding and joint decoding paradigm of DVC, enabling efficient compression of Wyner-Ziv (WZ) frames. At the encoder, a Variational Autoencoder (VAE) is employed to compress WZ frames into low-dimensional latent representations. At the decoder, a conditional diffusion model is introduced to guide the progressive denoising of a pure noise image, leveraging the latent representation of the WZ frame and the information from decoded Key frames to reconstruct high-quality WZ frames. Experimental results demonstrate that CD-DVC outperforms the traditional VVC/H.266 intra standard and the current deep learning (DL)-based end-to-end distributed video coding scheme, LDVC [1], in terms of reconstruction quality. This work provides a new solution for video compression in resource-constrained wireless environments. We open-sourced the implementation at GitHub<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">1</sup>.