Optimizing Video Quality Using Machine-Learning-Controlled Adaptive Resolution, Video Compression

数字视频需要大量的数据，并且在存储和/或传输之前必须进行压缩。电视广播公司、视频游戏制造商、视频内容分发商在其产品和/或服务中使用视频压缩。AMD为平板电脑、游戏机、嵌入式设备和云服务器设计和制造显卡和微处理器。这些产品必须产生尽可能好的视频质量。在视频压缩中，比特率和视频质量之间存在固有的权衡。因此，在给定的比特率下获得最佳的视频质量是一项至关重要的任务。通过在编码前降低视频序列的分辨率或在编码过程中使用较大的量化参数，可以降低视频序列的比特率。这两个选项中哪一个产生较少的质量损失取决于视频内容以及可用的网络带宽。本研究项目的目标是设计一种机器学习算法，以在运行时决定是否以原始高分辨率编码视频图片（或图片组），或者降低分辨率，以较小的量化步长编码较低分辨率的版本，在接收端解码和上采样，期望在给定的比特率下实现最佳质量。通过利用最佳分辨率/量化步长组合，我们开发的自适应视频分辨率调整方案可以导致显著的比特率节省目标质量或显著的质量改善目标比特率。

Digital video requires a huge volume of data and must be compressed before it can be stored and/ortransmitted. TV broadcasters, video game manufacturers, video content distributors use video compression intheir products and/or services. AMD designs and manufactures graphics cards and microprocessors for tablets,gaming consoles, embedded devices and cloud servers. These products must yield the best possible videoquality. In video compression, there is an inherent trade-off between bitrate and video quality. Obtaining thebest video quality for a given bitrate is, therefore, a crucial task. The bitrate of a video sequence can be reducedby reducing its resolution prior to encoding or by using a larger quantization parameter during encoding. Whichof these two options yields less quality loss depends on the video content as well as the available networkbandwidth. The goal of this research project is to design a machine learning algorithm to make a run-timedecision of whether to encode a video picture (or group of pictures) at the original high resolution, or to reduceresolution, encode the lower resolution version with a smaller quantization step, decode and upsample atreceiver side with expectation to achieve the best quality for a given bit rate. By utilizing the optimalresolution/quantization step combination, our developed adaptive video resolution adjustment scheme canresult in significant bitrate savings for a target quality or significant quality improvements for a target bitrate.