Video Super-Resolution Method Using Multi-Frame Attention and Gradual Fusion

The objective of video super-resolution is to increase the resolution of the video for providing good vision effect. The application of video super-resolution spreads over many areas such as movie, remote sensing, automatic driving, medical imaging, security system, manufacturing industry and military. A video super-resolution method based on multi-frame attention and gradual fusion is proposed. The multi-frame attention is used to capture temporal correlation between frames which can take advantage of the features in consecutive frames for reconstructing the current frame with more high resolution. Meanwhile, a gradual fusion is adopted to make feature fusion easier. We also proposed a low redundancy training strategy to accelerate the learning process of network and avoid repeat feature learning. The ablation experiment and the comparative experiment show that our method can improve the high frequency detail and the texture of video super-resolution without using any 3D convolution operations. In the former, the multi-frame attention is used to fuse the useful information of the adjacent frame to the objective frame to improve the high frequency details, in the latter; the gradual fusion and the low redundancy training strategy are applied to decrease the computation. The temporal feature extraction capabilities of the video super-resolution networks are enhanced to develop the high-resolution video better.