Personalized Semantic Feature Reconstruction Based Generative Portrait Video Compression

Traditional visual data (e.g., images and videos) compression focuses on signal fidelity to remove redundancy, while reducing semantic redundancy offers further potential to improve video compression performance. To this end, we propose a novel generative approach for portrait video compression using personalized semantic feature reconstruction. Specifically, we represent the semantic feature in portrait videos with personalized and common features. To achieve efficient feature extraction, we design a personalized feature extraction network that infers edge details, serving as a critical condition for generation. For realistic reconstruction, we introduce a two-stage feature fusion generative adversarial network (GAN). This network utilizes decoded personalized features for conditional generation while integrating common information like texture and color. Both qualitative and quantitative experimental results demonstrate that our method can achieve perceptually pleasing results at ultra-low bitrates with higher semantic fidelity.