Frame importance and temporal memory effect-based fast video quality assessment for user-generated content

User-generated content (UGC) has become increasingly popular, promoted by the widespread use of social media and mobile devices. Therefore, instant and immersive UGC video quality assessment is urgently needed to provide appropriate recommendations for video reviewers prior to distribution. However, existing methods are neither efficient at assessing UGC videos due to the expensive frame-by-frame process nor suitable for deployment on devices with limited computational capabilities because they require sophisticated GPU-dependent computation. In this paper, we propose a fast UGC video quality assessment method, named FastVQA, by considering both keyframe importance and human temporal memory effects. First, a novel key frame selection strategy based on feature entropy is developed to achieve efficient and accurate feature extraction. Inspired by human short-term and long-term memory effects, we design a temporal feature aggregation module by taking both local content details and global semantic information into consideration. Experimental results show that FastVQA can outperform the state-of-the-art (SOTA) methods on many datasets with significantly reduced CPU time, which implies that FastVQA can achieve a better balance between complexity and accuracy.