Research On Video Multicast Combined With Convolutional Neural Networks

At present,with the rapid development of wireless network,wireless video multicast technology has received widespread attention.In the compressed sensing theory,the measurements obtained by the observation array are not structurally distinguished.This advantage meets the requirements of wireless video multicast.Compressed sensing reconstruction combined with convolutional neural network can not only improve image reconstruction quality,but also reduce reconstruction time,we apply a combination of convolutional neural networks and compressed sensing to video multicast solutions,mainly completing the following two parts:Firstly,based on compressed sensing residual dense network for wireless video multicast(DCSRDN-Cast)is proposed.At the encoding,each frame of video is divided into non-overlapping blocks of fixed size 33×33,observe the image block using the observation matrix,and send the obtained observations to the noisy channel after packing.At the decoding,the user receives a data packet that matches his channel conditions,then unpacks and denoises it,then uses the trained network model to reconstruct it,and finally combines the image blocks into the final video image in order.The experimental results show that,undering the same channel noise ratio and packet loss rate,this solution greatly improves the video unicast and multicast performance than the video multicast solution based on deep compression-sensing fully connected network(DCSFCN-Cast).Secondly,we proposed based on the dilated convolution residual network for video multicast scheme(DRNDCS-Cast).This scheme aims to reduce the video reconstruction time without greatly affecting the quality of video reconstruction,and meet the needs of real-time video transmission.At the encoding,each frame of the video is divided into 33×33 image blocks,the observation values are obtained through observation array observation,and then the observation values are packaged and sent to the channel.At the decoding,the data packet is depacketized and denoised,and the trained network model isused to reconstruct it to obtain the final video image.The experimental results show that the video unicast and multicast performance is greatly improved over the DCSFCN-Cast solution,and the reconstruction time is shortened by about one third compared with the DCSRDN-Cast solution,which is more suitable for real-time transmission.