ERI: Generative Adversarial Networks for Video Coding

该奖项的全部或部分资金来自《2021年美国救援计划法案》(公法117-2)。视频编码是一项重要的技术，它对视频信号进行压缩，以节省传输带宽，并为互联网用户提供视觉上令人愉快的解码视频。受最近深度学习突破的启发，卷积神经网络已越来越多地被用于视频编码算法，以提供与传统方法相比显著的编码增益。然而，现有的基于卷积神经网络的视频编码方案往往会产生与人类感知不一致的模糊解码图像，并且这些方案的高计算复杂性阻碍了它们在智能手机和平板电脑等功率受限和计算资源受限的设备上的部署。最近，生成性对抗性网络展示了它在低比特率下解码清晰和逼真的图像的能力，但很少有人研究它在视频压缩方面的潜力。该项目将开发基于生成性对抗性网络的视频编码系统，以提高编码效率，同时提供高感知质量的解码视频。该项目还将研究低复杂度算法，以降低功耗并加快所提出的视频编码系统的推理速度，使其适用于移动和低延迟应用。该项目的成功有望加速流媒体视频服务的经济增长，造福人们的日常专业和娱乐活动。它将推进监控视频服务，以加强机场、办公室、高速公路和道路交叉口等场所的公共安全。该项目的研究活动将通过论文研究、高级设计项目以及机器学习和人工智能课程，为包括少数族裔和代表性不足群体在内的研究生和本科生提供培训机会。该项目的研究成果将在暑期工程研讨会项目中展示，以激励高中生攻读大学理工科专业。该项目将解决两个问题：(1)如何利用视频帧之间的时间相关性，并在基于生成性对抗性网络的视频编码体系结构中探索场景动态？提出了两种编码方法：分层预测编码方法和基于三维卷积的空时编码结构。由于现有的生成性对抗网络模型大多是用于静态图像压缩，本研究的成功将为视频编码专业人员打开基于生成性对抗网络的编码系统的大门。(2)如何降低深度视频编码网络的计算复杂度？尽管基于深度学习的视频编码工具具有性能优势，但目前在真实场景中采用的视频编码工具很少。这是由于计算复杂性高、推理速度慢以及与深度网络计算相关的大图形处理单元内存需求。为了解决这个问题，提出的研究将通过可分离的卷积运算来开发算法来降低基于深度学习的视频编码模型的复杂度、模型规模和模型参数。这项研究成果将加速深度视频编码模型在现实世界应用中的部署。这一奖项反映了NSF的法定使命，并通过使用基金会的智力优势和更广泛的影响审查标准进行评估，被认为值得支持。

This award is funded in whole or in part under the American Rescue Plan Act of 2021 (Public Law 117-2).Video coding is an important technology that compresses video signals to save transmission bandwidth and to provide Internet users with visually pleasing decoded videos. Inspired by recent breakthroughs in deep learning, convolutional neural networks have been increasingly exploited into video coding algorithms to provide significant coding gains compared to conventional approaches. Nevertheless, existing convolutional neural network-based video coding schemes tend to generate blurry decoded images which are inconsistent with human perception, and the high computational complexity of these schemes hinders their deployment on power-constrained and computation resource-limited devices, such as smart phones and tablets. Recently, the generative adversarial network demonstrated its capability of decoding sharp and photo-realistic images at low bit rates, but little research has investigated its potential for video compression. This project will develop generative adversarial network-based video coding systems to enhance the coding efficiency, meanwhile providing decoded videos with high perceptual quality. The project will also investigate low-complexity algorithms to reduce the power consumption and to accelerate the inference speed of the proposed video coding systems so that they are suitable for mobile and low-latency applications. The success of the project is expected to accelerate the economic growth of streaming video services to benefit people’s daily professional and entertainment activities. It will advance surveillance video services to enhance public safety in places such as airport, offices, highway, and road intersections. The research activities of the project will provide opportunities to train graduate and undergraduate students including minority and under-represented groups through theses research, senior design projects, as well as machine learning and artificial intelligence courses. The research results of the project will be showcased in a summer engineering seminar program to motivate high school students to pursue science and engineering majors in college.This project will address two problems: (1) How to leverage temporal correlations among video frames and explore scene dynamics in a generative adversarial network-based video coding architecture? Two approaches are proposed: a hierarchical predictive coding approach, and a spatial-temporal coding architecture based on 3-dimensional convolution. Since most existing generative adversarial network models are for still image compression, the success of this research will open the door to generative adversarial network-based coding systems for video coding professionals. (2) How to reduce the computational complexity of deep video coding networks? Despite the performance benefits of deep learning-based video coding tools, few of them are currently being adopted in real-world scenarios. This is due to the high computational complexity, slow inference speed and the large graphic processing unit memory requirements associated with deep network computation. To address this problem, the proposed research will develop algorithms to reduce the complexity, model size and model parameters of deep learning-based video coding models via separable convolution operations. The research results will accelerate the deployment of deep video coding models in real-world applications.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.