深度学习实时视频超分辨率重建实验设计

Experiment Design of Real-time Video Super-resolution Reconstruction Based on Deep Learning

针对现有视频超分辨率重建深度学习网络结构复杂、实时性差的实验问题,设计了基于深度学习的实时视频超分辨率重建实验。提出将Ghost-Module结构应用于循环递归神经网络,并借助改进的残差学习结构加速网络收敛,实现了视频超分辨率的实时重建。实验结果表明:新算法的运算速度为FRVSR算法的5.4倍;在参数量与TOFlow算法相近的基础下,运行速度是TOFlow模型的34.5倍,是高性能算法DUF-52L的28.8倍;针对480×270低分辨率视频影像可实现24 f/s的实时4倍超分辨率重建。所设计的轻量级循环卷积视频超分辨率重建网络在保证视频超分辨率重建质量的基础上极大地提高了重建速度。

Aiming at the problem of complex structure and long processing time in traditional video super-resolution reconstruction algorithm,this paper proposes a new algorithm,in which the GhostModule structure is applied in the recurrent recursive neural network and the residual structure is used to accelerate the convergence of the network.Experimental results show that the speed of the new algorithm is 5.4 times faster than that of FRVSR algorithm;the parameters of the new algorithm are similar to those of TOFlow algorithm,but the running speed is 34.5 times;the speed of new algorithm is 28.8 times faster than that of DUF-52L.It realizes the real-time 4x super-resolution reconstruction with the speed of 24 frames per second for the low-resolution video with resolution of 480×270.The proposed lightweight recurrent convolution video super-resolution reconstruction network not only has good reconstruction quality,but also greatly improves the reconstruction speed.