Compression-RQ-VQA: Leveraging Rich Quality-Aware Features for Compressed Video Quality Assessment

Abstract

Video compression technology has become an indispensable part of modern video transmission systems. The main challenge in blind (no-reference) quality assessment of compressed videos lies in the complex perceptual quality differences caused by variations in codec systems, as well as the various inherent quality and content of the videos. In response to this challenge, we propose a conceptually simple yet highly effective model in this paper, the Compression - R ich perceptual Q uality aware V ideo Q uality A ssessment model. The model is based on a feature ensemble strategy, comprising two main components. The main branch employs the Simple-VQA model, pre-trained on user-generated content (UGC) video quality assessment tasks. The backup feature branch extracts features using the LIQE model (focusing on video semantics and distortion type perception); the Q-Align model (based on large multi-modal models (LMMs)); the FAST-VQA model (based on block-level local quality perception); and the Compression-Net model (sensitive to local compression perception). By integrating these rich quality perception features, the model’s performance in evaluating the quality of compressed videos is significantly enhanced. The proposed model has secured 2nd place in the ECCV-AIM-2024 Compressed Video Quality Assessment Challenge . Moreover, experimental results demonstrate that the proposed model performs best on four public VQA datasets.