Conformer Based No-Reference Quality Assessment for UGC Video

Due to the lack of reference videos and complex distortions, video quality assessment (VQA) for user-generated content (UGC) has become a highly challenging task. Previous studies have fully demonstrated the effectiveness of deep learning models in UGC VQA, but most methods only use a single CNN or Transformer to extract features, without fully integrating the advantages of them. In this paper, we propose a no-reference video quality assessment method based on Conformer, which utilizes convolutional neural networks and self-attention mechanisms in parallel to obtain features more suitable for UGC video quality. A Feature Attention (FA) module is further proposed to help the model focus on important parts of the video. The experimental results show that the proposed model achieves good performance on mainstream subjective UGC video quality databases, indicating the effectiveness on UGC VQA.