多注意力结合光流的视频超分辨方法

针对现如今监控摄像、卫星遥感以及视频娱乐等领域对视频图像的清晰度要求越来越高,而目 前大部分视频超分辨方法存在参数量大、恢复的视频存在抖动等问题,提出了一种多注意力结合光流的视频 超分辨方法,通过引入多个注意力包括空间注意力、通道注意力以及自注意力来提升超分辨性能。 具体而 言,作为一种特征加权的增强方法,这些注意力可以捕获视频帧的时空特征并增强自适应性和通道间的依赖 性,实现全局学习的功能;同时,提出双阶段特征对齐思路,首先利用光流对视频进行估计,进行第一阶段的 特征对齐,然后引入长短是记忆网络结构增强位置和通道的特征融合,进行第二阶段的特征对齐,以防止恢 复的视频帧出现抖动。 实验结果表明:该方法在评估标准和可视化效果方面都取得了令人满意的效果。

Video Super-resolution Method of Multi-attention Combined with Optical Flow

In view of the increasingly higher requirements for the definition of video images in the fields of surveillance cameras, satellite remote sensing, and video entertainment, most of the current video super-resolution methods have problems such as large parameters and jitter in the restored video. A video super-resolution method combining multiple attention and optical flow is proposed. By introducing multiple attentions including spatial attention, channel attention and self-attention, the super-resolution performance is improved. Specifically, as a feature-weighted enhancement method, these attentions can capture the spatio-temporal features of video frames and enhance the adaptability and inter-channel dependence to achieve the function of global learning. At the same time, the idea of two-stage feature alignment is proposed. First, the optical flow is used to estimate the video, the first stage of feature alignment is performed, the introduction of length is a feature fusion of memory network structures that enhance locations and channels, and the second stage of feature alignment is performed to prevent the recovered video frame from shaking. The experimental results show that the method has satisfactory results in both the evaluation standard and the visualization effect.