Deblurring Method Of Traffic Video Based On Depthwise Separable Residual Network

As an important information medium,the definition of video and image seriously affects people’s sensory experience.When the camera shakes or vehicle speed is too fast,it is very easy to appear motion blur in collected traffic video and image.The motion blur will affect the performance of intelligent transportation system,and even threaten the safety of driverless.At present,the deblurring method of video and image based on residual network has achieved good restoration effect,but there are still some problems such as complex model,more parameters,insufficient feature extraction process.Aiming at these problems,this paper improves the residual network,proposes image and video deblurring algorithm,and applies the algorithm to traffic video deblurring task.The main contents of this paper is as follows:1.An image deblurring network model based on depthwise separable residual network is proposed,which consists of feature extraction network and upsampling reconstruction network.In the feature extraction network,the depth separable convolution is introduced to replace the convolution method in the original residual network.At the same time,the BN layer is removed and the dropout layer is added to prevent over fitting.The improved residual network is combined with the feature pyramid and attention mechanism to achieve multi-scale feature extraction and fusion,which can reduce the amount of model parameters and achieve better deblurring effect.The test results show that the Peak signal-to-noise Ratio(PSNR)and Structural Similarity(SSIM)of improved model are 31.65 d B and 0.936 respectively,which are 4.59% and 0.43% higher than SRNDeblur proposed by Tao et al.2.A model based on discontinuous frame alignment and fusion for video deblurring is proposed.In view of the fact that the current video deblurring methods do not fully consider the spatio-temporal relationship between the frames of video,this paper proposes an discontinuous frame alignment and fusion method which aligns the frames at the feature level according to the way of discontinuous frame,and fuses them based on spatio-temporal attention mechanism to improve the effect of alignment fusion.The video deblurring model combine the discontinuous alignment fusion method and image deblurring model,and experimental results show that the PSNR and SSIM of the model are 32.51 d B and 0.943 respectively,which are 1.18% and 1.75% higher than CDVD-TSP that proposed by Pan et al for video deblurring.3.The video deblurring model proposed in this paper is applied to traffic video deblurring task,which is tested in different time,different weather and different types of roads.The test results show that the deblurring task could be well completed in most scenes.