Multi-Frame Quality Recovery Model for Compressed Video Enhancement

Video quality is essential for many consumer electronic devices. However, existing quality enhancement methods do not pay enough attention to the video characteristics. The success of artificial intelligence (AI) can help improve the quality of compressed video. In this paper, we perform an analysis of compressed video. Based on this, we propose utilizing multi-frame information as input and using two adjacent high-quality frames to enhance the low-quality frames in between. Therefore, an efficient Multi-Frame Quality Recovery (MFQR) method is proposed. In MFQR, we first develop a forward feature extraction module to extract information from multi-frame input. Then, a bidirectional time information extraction module, consisting of a Bi-directional Long-Short Time Memory (Bi-LSTM) structure and channel attention mechanism, is developed to capture bi-directional temporal information and enhance spatiotemporal information expression of features. Finally, we construct a residual feature enhancement module to improve model performance. Extensive experiment results show that the proposed MFQR method achieves an average increase of 27% in PSNR and reduces the number of parameters by an average of 23% than the representative methods.