面向视频去模糊的高效时空特征提取循环神经网络

Efficient Spatio-Temporal Feature Extraction Recurrent Neural Network for Video Deblurring

针对现有基于循环神经网络的视频去模糊算法跨帧信息聚合能力有限、计算效率低的问题,提出高效时空特征提取循环神经网络的算法.首先结合残差密集网络结构和通道注意力机制,从给定序列图像的各帧中高效提取并选择表达力强的特征类型;然后提出循环神经网络框架下的时空特征增强融合模块,从高度冗余和干扰性强的序列图像中,统筹选择有效信息并将其融入到当前帧特征图中,补充当前帧信息;最后通过重建模块将增强后的特征图转换为去模糊后的当前帧图像.在3个包含合成和真实模糊视频数据的公开数据集上的定量和定性实验结果表明,所提算法能够以较小的计算成本取得优异的视频去模糊效果,在GOPRO数据集上,PSNR达到31.43 dB,SSIM达到0.9201.

Considering that existing recurrent neural network-based video deblurring methods are limited in cross-frame feature aggregation and computational efficiency,an efficient spatio-temporal feature extraction recurrent neural network is proposed.Firstly,we combine a residual dense module with the channel attention mechanism to efficiently extract discriminative features from each frame of a given sequence.Then,a spatio-temporal feature enhancement and fusion module is proposed to select features from the highly redundant and interfering sequential features and integrate them into the features of the current frame.Finally,the enhanced features of the current frame are converted into the deblurred image by a reconstruction module.The quantitative and qualitative experimental results on three public datasets,containing both synthetic and real blurred videos,show that the proposed network can achieve excellent video deblurring effect with less computational cost.Among them,on the GOPRO dataset,the PSNR reaches 31.43 dB and the SSIM reaches 0.9201.