| The requirements of ultra-high panoramic videos pose a challenge to the transmission technology for panoramic videos.After a panoramic video is projected from a sphere to a twodimensional grid plane,stream files to be transmitted are generated by compression.Accordingly,the design of a transmission strategy needs to consider the influence of projection format.In order to provide users with an immersive experience,the spatial resolution of ultrahigh panoramic videos is extremely high,which requires huge bandwidth resources.Tile-based transmission scheme for panoramic videos requires adaptive scheduling that considers both space and quality to achieve the optimal users’ experience in a bandwidth-restricted environment.For the core problems of transmission optimization for panoramic videos,this paper proposes three solutions based on deep reinforcement learning in the aspects of content production and transmission process.(1)The Cubemap projection is often used to project panoramic videos.However,the default Cubemap projection causes easily some foreground objects with obvious distortion to cross multiple projection planes without considering the distribution of the objects.To solve the above problem,this paper proposes a content-aware cubemap projection for panoramic image based on Deep Q-learning(DQL)that quickly predicts the horizontal rotation angle in a limited number of iterations by analyzing the distribution of foreground objects for optimizing the Cubemap projection.Experimental results show that,compared with existing optimization methods for the Cubemap projection,the method proposed effectively reduces the distortion of foreground objects and improves the compression rate of panoramic images.(2)For the optimization of the transmission process for panoramic videos,this paper proposes a transmission scheme for panoramic videos applied to the single-user on-demand scene,which adopts the cooperative transmission method of the basic layer and the enhancement layer to transmit high-bitrate tiles in the field-of-view(Fo V)and low-bitrate basic chunks for covering area outside the Fo V.During execution,firstly the Fo V prediction algorithm is used to predict accurately the Fo V in the next few seconds,known as the predicted Fo V;Then the adaptive bitrate algorithm based on Asynchronous Advantage Actor-Critic(A3C)is used to select the bitrate of each tile in the predicted Fo V and dynamically reduce the prefetch time of tiles for improving the accuracy of the Fo V prediction algorithm.Experimental results prove that in a single-user scenario,the scheme can obtain higher quality of experience during the entire transmission on demand.(3)In order to simultaneously optimize the viewing experience of multiple users in a multiuser on-demand scenario,based on the above scheme,this paper proposes a transmission scheme for panoramic videos applied to the multi-user on-demand scene,whose core is to add edge computing and edge cache to the edge device(such as a base station)of the content distribution network.The base station uses the request queue to collect and organize the video requests of different users,analyzes factors such as bandwidth and edge cache,and uses the A3C-based global bitrate algorithm to select the bitrate of each tile in multiple Fo Vs.The edge cache is used to store frequently used tiles in a short time,which is convenient for repeated use.In order to further efficiently use the edge cache,this paper proposes an optimized replacement algorithm related to the Fo V,which effectively improves the efficiency of tile replacement.Experimental results prove that in the multiuser scenario,this scheme can provide a high-quality immersive experience for all users and reduce the throughput of the server without changing the downlink bandwidth of the server. |
