A Video Deblurring Method Based On Spatiotemporal Feature Learning And Fourier Aggregation

In the field of video shooting,poor image acquisition environment,degradation of the imaging system and target movement often lead to undesired blur in captured video.Researchers have proposed many excellent video deblurring methods.However,due to the unique spatio-temporal correlation of video images,there is still a lot of room for improvement in utilizing spatiotemporal information comprehensively.This paper reviews the research status at home and abroad,analyzes and implements three classical video deblurring algorithms: weighted Fourier accumulation algorithm(WFA),deep video deblurring network algorithm(DBN)and spatio-temporal recurrent network algorithm(STRCNN).In view of the shortcomings of the above algorithms,the following works have been done in this paper:This paper proposes a video deblurring algorithm based on structured confidence incremental weighted Fourier accumulation.Weighted Fourier accumulation algorithm are used to calculate the weights of all frames in local temporal window,which lacks adaptive weighting mechanism of quality control.In our algorithm,the sharpness of the local video window image is sorted based on the structure confidence,and the restoration frame is reconstructed by conbining the clearest image in the ordered window with the deblurred image generated in the last iteration.The further iteration is judged according to the reconstruction image quality in the way of incremental control,so as to effectively avoid the adverse effect of over blurring the adjacent frames.Experimental results prove that the deblurring performance of this algorithm is significantly improved compared to the original weighted Fourier aggregation algorithm.This paper proposes an effective Fourier accumulation embedded 3D convolutional encoder-decoder network for video deblurring.In order to solve the problem that the traditional deep learning video deblurring algorithm can't make full use of the temporal correlation of video,a synthetic network of 3D convolutional neural network and Fourier aggregation is designed.Specifically,this algorithm learns the spatiotemporal features of video frames in local window through the 3D convolutional network.The intermediate deblurring image comes from Fourier transformation,and the Fourier accumulation module is used to further learn the Fourier features of the image.The deblurring video is finally generated according to the trained depth network.Experimental results prove that this method has a good performance of deblurring with fast processing speed.Finally,this paper integrates the above algorithms and develops a jitter video deblurring software.Due to the fact that jitter and blur often appear in degraded video at the same time,the system integrates an excellent video stabilization method,realizes the joint processing of video jitter and blur degradation.In order to visually compare the restoration effects of different algorithms,the system also provides the functions of reconstruction video visualization and video quality evaluation...