The Reasearch Of Deep Learning Based For Panoramic Video Encoding Post-processing

Panoramic video becomes more and more universal in people's life.However,due to the limited performance of network transmission,high-resolution panoramic video will be compressed by mature coding methods such as HEVC,which seriously affects the visual experience of users.Therefore,how to reasonably post-process the compressed panoramic video becomes a necessary problem,which is the primary content of research in this paper.In this paper,according to different encoding Settings(I frame and P frame)and combined with the characteristics of panoramic video,two different panoramic video post-processing algorithms are proposed to adapt to different encoding strategies of panoramic video post-processing.The primary research methods and innovative results of this paper are as follows:1)ARGAN,a panoramic video post-encoding algorithm,is proposed based on single frame information to adapt to I-frame encoding.ARGAN adopts GAN(Generative Adversarial Networks)design,make the Generative Adversarial Network learn to form the way that uncompressed video is constructed,combined with perception loss to increase visual authenticity post-processing video frame.ARGAN can effectively remove artifacts and block effects while achieving good visual effects.Meanwhile,ARRESNET,a separate generator network in ARGAN,can achieve superior PSNR and SSIM promotion effects(2)A panoramic video post-encoding algorithm ATT-RMDF is proposed based on multi-frame information to adapt to P-frame encoding.This algorithm fuses the information of multiple frames for post-processing of target frames and can obviously optimize the fluctuation of video frame quality caused by encoding.Avoiding explicit optical flow prediction and using deformable convolution to alignment multi-frame information,the attention residuals block was designed to greatly improve the model effect by using residuals learning,while enabling the network to make targeted use of the information of each adjacent frame,and the Align Val Net is designed to supervise the accuracy of alignment.Experimental results prove that the ARGAN based on single frame proposed in this paper can reach a very realistic visual effect while removing the compression artifact and block effect,and ARRESNET,the independent generator network of ARGAN,has a high PSNR/SSIM repair effect.The ATT-RMDF based on multiple frames can not only effectively recover PSNR and SSIM,but also obviously repair the quality fluctuation before the video frames caused by encoding.