Research On Video Frame Interpolation Based On Deep Learning

Video frame interpolation technology is to synthesize one or multiple frames between two consecutive video frames for frame rate conversion.This technology can solve two problems:(1)to increase the video frame rate to improve people’s perception;(2)in the case of limited network bandwidth,by first transmitting the videos of low frame rate and then synthesizing the videos of high frame rate at the receiving end,to reduce the burden of video transmission.The steps of traditional video frame interpolation methods,which are based on decoding motion vector,block matching and optical flow,are motion estimation and pixel synthesis,but these methods often fail to achieve good results.With the development of deep learning,researchers use kernel-based and optical flow-based convolutional neural networks for video frame interpolation.The performance of these methods is much better than that of the traditional ones.In recent years,generative adversarial networks have achieved good results in the field of image generation.In this thesis,we use generative adversarial networks to optimize convolutional neural networks.The innovations are as follows:1.We propose a generative adversarial network which uses the kernelbased convolutional neural network as the generator and the discriminator to discriminate the frames generated by the generator.Through adversarial training,the intermediate frames are generated continuously.The similarity between the optimal intermediate frames and the original ones is calculated to measure the performance of the model.2.We propose a generative adversarial network which uses the optical flow-based convolutional neural network as the generator and the discriminator to discriminate the frames generated by the generator.Through adversarial training,the intermediate frames are generated continuously.The similarity between the optimal intermediate frames and the original ones is calculated to measure the performance of the model.In this thesis,we also design a visual system for displaying the result.It splits the video of low frame rate into a frame sequence,and uses the above models to interpolate frames to synthesize the video of doubled frame rate.The properties of the two videos prove that the model proposed in this thesis can realize the frame rate conversion and improve the smoothness of video.