基于通道注意力机制的视频超分辨率方法

A video super‑resolution method based on channel attention

基于视频帧间信息特征,提出了基于通道注意力机制的循环残差注意力网络,将连续的低分辨率视频帧、前一时刻输出帧和隐藏态作为输入进行特征提取,在隐藏态中引入残差连接和注意力机制,增强网络特征提取能力,经过亚像素卷积层重建出高分辨率视频帧。然后将本视频超分辨率网络模型在Vid4、UDM10、SPMCS视频数据集进行了测试。实验结果表明,与其他基于深度学习的视频超分辨率方法相比,本方法能利用帧间特征信息较好地恢复高频特征信息,恢复的视频图像PSNR和SSIM值都比其他主流方法要高,同时取得了较好的主观视觉效果。

Based on the information between video frames,a recurrent residual attention network based on channel attention mechanism was proposed in this paper.Continuous low‑resolution video frames,along with the output frame and the hidden state from the previous moment were input into the network for feature extraction.The residual connection and the attention mechanism were introduced in the hidden state to enhance the network′s feature extraction ability.The high‑resolution video frames were reconstructed through the sub‑pixel convolution layer.The video super‑resolution network model presented in this paper was tested on Vid4,UDM10 and SPMCS video datasets.The experimental results show that compared with other deep learning based video super‑resolution methods,the proposed method uses interframe feature information to better recover high‑frequency feature information.The PSNR and SSIM values of the recovered video images outperform other general methods,while the subjective visual effect is better.