基于三维时空注意的密集连接视频超分算法

Densely connected video super-resolution based on three-dimensional spatial-sequential attention

针对视频超分对时间帧间信息以及分层信息的利用不充分,设计了一种具有空间时序注意力机制的密集可变形视频超分辨率重建网络。利用三维卷积来提取经可变形卷积模块对齐后的相邻帧之间的时间序列信息,同时设计具有步幅卷积层的轻量级模块来提取空间注意力信息。在特征重构阶段引入密集连接,充分利用分层特征信息以实现更好的特征重建。选取公共数据集进行实验验证,结果表明,提出的算法在客观评价指标与视觉对比效果上都有提升。。

Aiming at the insufficient utilization of temporal inter-frame information and hierarchical information in video super-resolution,a dense deformable video super-resolution reconstruction network with spatial-sequential attention mecha-nism is designed.Three-dimensional convolution is used to extract sequence information between adjacent frames aligned by deformable convolution module,and a lightweight module with strided convolution layer is designed to extract spatial attention information.Dense connections are introduced in the feature reconstruction stage to make full use of hierarchical feature information to achieve better feature reconstruction.The public datasets are selected for experimental verification.The results show that the proposed algorithm has improved both objective evaluation indicators and visual contrast effects.