基于运动特征融合的快速视频超分辨率重构方法

基于深度学习的视频超分辨率重构方法常面临重构精度不高或重构时间过长的问题,难以实时获得高精度的重构结果.针对此问题,文中提出基于深度残差网络的视频超分辨率重构方法,可以快速地对视频进行高精度重构,并在较小分辨率视频的重构过程中达到实时重构的要求.自适应关键帧判别子网自适应地从视频帧中判别关键帧,关键帧经过高精度关键帧重构子网进行重构.对于非关键帧,将其特征与邻近关键帧间的运动估计特征和邻近关键帧的特征逐层融合,直接获得非关键帧的特征,从而快速获得非关键帧的重构结果.在公开数据集上的实验表明,文中方法能实现对视频的快速、高精度重构,鲁棒性较好.

Fast Video Super-Resolution Reconstruction Method Based on Motion Feature Fusion

Video super-resolution reconstruction methods based on deep learning are often faced with the problems of long time consumption or low accuracy. A video super-resolution reconstruction method based on deep residual network is proposed. It reconstructs videos with high accuracy quickly and meets the real-time requirements for low-resolution videos. Firstly, the adaptive key frame discrimination subnet is utilized to adaptively identify key frames from the video. Then, the reconstruction results of the key frames are obtained by the high precision reconstruction subnet. For non-key frames, the reconstruction results are directly gained based on the features obtained by fusing the features of the corresponding key frame and the motion estimation features between the non-key frame and the adjacent key frame. Experiments on open datasets show that videos are fast reconstructed by the proposed method with high accuracy and robustness.