Research On Deep Learning Based Video Frame Interpolation Algorithm

As the current mainstream form of information transmission,video is of great significance in social life.The reduction in bandwidth caused by objective factors of information channels affects the rate of video transmission from the source to the destination,which in turn decline user's experience of video.Transmitting low frame rate video could compress the data volume,but there will be a visual freeze when the users are watching the video.Video frame interpolation algorithm can well solve the visual problem.In the meanwhile,video frame interpolation algorithm can also be used to solve other problems such as generating slow-motion video,and is important in daily life.In this paper we focus on video frame interpolation algorithm based on deep learning and proposed our method to improve the performance of existing work.In this paper a video frame interpolation algorithm was proposed based on featureprior modeling.To address the problem of high-speed object optical flow calculation failure during video frame interpolation,a model that have priority on modeling of object's features,which increased the influence of key points in images on output optical flow,was designed and verified with experiments.The model improved performance of huge-motion optical flow calculating.By comparison with traditional methods and mainstream deep learning methods,our model showed better performance.Specially,our model(SuperSloMo+feature-prior)has a 0.32dB improvement in PSNR and the SSIM increase about 0.005,with IE reduced by 0.25/pixel,compared to model that didn't use our method.In order to further optimize the model,we proposed a video frame interpolation algorithm based on joint optimization.Based on feature-prior model we proposed,the feature optical flow calculation network is multiplexed to reduce network parameters,and a super-resolution network is used to jointly optimize our model.In the experiment,we verified that joint optimization can improve the performance of feature-first model,about 1.34dB on PSNR,0.009 on SSIM and 0.96/pixel on IE.