A recurrent video quality enhancement framework with multi-granularity frame-fusion and frame difference based attention

Abstract

In recent years, deep learning has attracted substantial research attention for video restoration. Among the existing contributions, the single-frame based approaches purely rely on one reference frame and neglect the rest neighboring frames when enhancing a target frame. By contrast, the multi-frame based contributions exploit temporal information in a sliding window and the existing recurrent design only employ a single preceding enhanced frame. It is intuitive to exploit both multiple original neighboring frames and the preceding enhanced frames for video quality enhancement. In this paper, we propose a Recurrent video quality Enhancement framework with Multi-granularity frame-fusion and frame Difference based attention (REMD). Firstly, we devise a three-dimensional convolutional neural network based encoder-decoder fusion model, which fuses multiple frames in multi-granularity. Secondly, severe compression artifacts tend to emerge on the edges and textures of the compressed frames. We propose a frame difference based spatial attention method to intensify the edges and textures of motioning regions. Finally, a recurrent sliding window design is conceived for exploiting the temporal information in preceding enhanced frames and subsequent neighboring frames. Experiments demonstrate that our method achieves superior performance in comparison to the state-of-the-art contributions with substantially reduced spatial and computational complexity.