Video Interpolation Technology Based On Deep Learning

Video frame interpolation is the application of computer vision technology in the field of video enhancement.The goal is to increase the frame rate of the video and enhance the continuity of visual effects.Benefiting from the rapid development of deep learning technology,video interpolation technology based on convolutional neural networks has become an important development direction in recent years.At present,a series of methods combining optical flow estimation and deep learning represent the advanced level in the field of video interpolation.However,there are many complex scenes in real scenes.For example,the movement of small objects is challenging for optical flow estimation and video interpolation.In order to enhance the ability of model motion estimation and further improve the effect of video frame interpolation,the optical flow model and frame synthesis model in this paper are proposed.The main innovations of this article are as follows:(1)By improving the instance normalization,a reversible instance normalization is proposed.Aiming at the problem that the hardware equipment of this paper is difficult to use enough batch size to support the effect of batch normalization,this paper improves the problem that the instance normalization is easy to lose information,which is used to replace the batch normalization.(2)An adaptive dense warping layer is proposed to replace replace the backward warping and cost volume structure in the current optical flow network.By clarifying the positional relationship between the matching point and the initial optical flow mapping point,the uncertainty for a single pixel is reduced,and the efficiency of searching and matching the optical flow network is effectively improved.(3)The feature attention mechanism is introduced,and the grouping error is proposed,which effectively reduces the effective information lost during the dimensionality reduction of the matching point features in the search and matching process,and facilitates the accuracy of the convolutional neural network to distinguish different matching points.(4)Warping the image features at multiple scales,reducing the loss of original image information during the forward deformation process,and providing effective information for the convolutional neural network to restore the intermediate frames,thereby improving the effect of video interpolation.