Research And Application Of Video Super-resolution Based On Convolutional Neural Network

Convolutional neural network(CNN)is a special type of depth neural network(DNN),and CNN has been successfully applied to image or video super-resolution,as well other restoration and detection tasks.The method in this paper is to extract the feature of interesting content of continuous sequence by depth convolution neural network.The extracted features are trained by depth convolution neural networks to generate learning models.In the process of recovering low resolution image frames,a priori knowledge obtained from the learning model is introduces to obtain high frequency details of the image frames.In the third chapter of this thesis,a video super-resolution algorithm based on convolution neural network is proposed,which extends to super resolution based on image super-resolution algorithm.Video super-resolution algorithm is a super resolution reconstruction method based on content clustering,which is consistent with video sequences in the interesting spatial domain and complex time domain,and the use of motion compensation and pre-training method to reduce the training time and improve the reconstructed video quality.In the recovery process,first of all to each data set between adjacent frames of motion modeling and estimation,comprising a plurality of adjacent frames in the training process,get additional information between each frame sub pixel motion obtained,then additional information through these frames difference can also be captured based on learning method.Then,the content is extracted,the content is clustered,or the content of interest is called "frame",the results show that the image and video super-resolution algorithm proposed in this thesis can better improve the quality and performance of image and video continuous frames.