Abstract
Adaptive streaming has been commonly used in content distribution and has become a primary tool to improve user experience performance. A video is encoded into multiple quality levels in adaptive streaming and divided into defined chunks of volume. The user then executes an adaptive bitrate algorithm to pick a suitable value for the specified network bandwidth. Nonetheless, there is a fundamental drawback that video quality depends heavily on the available server-client bandwidth. Therefore, video quality suffers directly when the network is congested. In current streaming of content (adaptive streaming), as network is congested, video quality degrades. For this job, in addition to traditional video codecs, we apply super-resolution Deep Neural Networks by adding performance enhancements frame by frame. Next, the DNN processor decodes a chunk of video into images. Then the DNN processor uses the DNN super resolution. The resulting frames are then re-encoded to video chunks that replace the playback buffer's original chunks. Just like in current adaptive streaming, an ABR algorithm is run by a device to select a proper video resolution. Instead, after viewing a low-resolution video ranging from 240p to 720p, the consumer actively uses its computing power to apply a super-resolution DNN that produces a 1080p or higher-resolution video. Neural Architecture Search on Video Super Resolution is introduced to perform super-resolution over existing methods to produce a faster and lighter model of weight that delivers better performance.