Image and Video Compression Meets Computer Vision

图像和视频压缩与计算机视觉的结合

• 批准号: RGPIN-2020-04525
• 负责人: Liang, Jie
• 金额: $ 3.35万
• 依托单位: Simon Fraser University
• 依托单位国家: 加拿大
• 项目类别: Discovery Grants Program - Individual
• 财政年份: 2021
• 资助国家: 加拿大
• 起止时间: 2021-01-01 至 2022-12-31
• 项目状态: 已结题
• 来源: https://www.nserc-crsng.gc.ca/ase-oro/Details-Detailles_eng.asp?id=740017
• 关键词: Image; Video; Compression; Meets; Computer

Recently tremendous progresses have been made in the theories of artificial intelligence, thanks to the advent of deep learning, a branch of machine learning theory that has been successfully applied in many fields and created numerous new business opportunities. In this program, we focus on the applications of deep learning in image/video compression and computer vision. Traditionally, image/video compression was mainly studied using signal processing, with little understanding of the contents, whereas computer vision aimed at understanding the contents using machine learning techniques. Since deep learning has become a powerful tool for both image/video compression and computer vision, it is time to unify the two fields more closely. Our long-term goal is to bridge the gap between image/video compression and computer vision. In the short term, we aim to develop new algorithms in the following topics: 1) deep learning-based image/video compression. 2) Compression/privacy-friendly designs for various computer vision tasks. The former will allow image/video compression schemes to achieve better compression of the images/videos by analyzing the contents of the images/videos. The latter will jointly design the architectures of the compression and computer vision to achieve better performance, and will ensure users' privacy in computer vision applications. The techniques we will employ in this program include advanced signal processing and machine learning tools such as trellis-coded quantization, range batch normalization, octave convolution, and federated learning. Video traffic currently accounts for about 80% of all internet traffic. As 5G mobile networks start to be deployed, more video traffics will be enabled. 5G will also significantly accelerate the adoption of various intelligent internet-of-things applications. By applying the theories and insights in image/video compression and computer vision to each other, we can obtain novel solutions with better performance, and create important paradigm shifts in both fields. It can also cultivate some innovative and disruptive applications that involve compression, computer vision, and deep learning. The privacy-preserving intelligent solutions developed in this program and their practical applications will make our life safer and more convenient without worrying about sacrificing our privacy. We also aim to develop fast and high-performance solutions that can be used in mobile and internet-of-things scenarios. This program will provide solid theoretical foundations and practical programming skills to the students involved, making them well prepared for their future careers in both image/video processing and computer vision. They will make valuable contributions to Canada's economy and help Canada to maintain its leadership position in the world in artificial intelligence and its applications.

近年来，由于深度学习的出现，人工智能理论取得了巨大的进展，深度学习是机器学习理论的一个分支，已成功应用于许多领域，并创造了许多新的商业机会。 在这个项目中，我们专注于深度学习在图像/视频压缩和计算机视觉中的应用。传统上，图像/视频压缩主要是使用信号处理来研究的，对内容的理解很少，而计算机视觉旨在使用机器学习技术来理解内容。由于深度学习已经成为图像/视频压缩和计算机视觉的强大工具，现在是时候更紧密地统一这两个领域了。我们的长期目标是弥合图像/视频压缩和计算机视觉之间的差距。短期内，我们的目标是在以下主题中开发新算法：1）基于深度学习的图像/视频压缩。2)适用于各种计算机视觉任务的压缩/隐私友好设计。前者将允许图像/视频压缩方案通过分析图像/视频的内容来实现图像/视频的更好压缩。后者将联合设计压缩和计算机视觉的架构，以实现更好的性能，并将确保用户在计算机视觉应用中的隐私。我们将在此计划中使用的技术包括高级信号处理和机器学习工具，如网格编码量化，范围批量归一化，倍频程卷积和联邦学习。 目前，视频流量约占所有互联网流量的80%。随着5G移动的网络开始部署，更多的视频流量将被启用。5G还将大大加快各种智能物联网应用的采用。通过将图像/视频压缩和计算机视觉的理论和见解相互应用，我们可以获得具有更好性能的新颖解决方案，并在这两个领域创造重要的范式转变。它还可以培养一些创新和颠覆性的应用程序，涉及压缩，计算机视觉和深度学习。该计划开发的隐私保护智能解决方案及其实际应用将使我们的生活更安全，更方便，而不必担心牺牲我们的隐私。我们还致力于开发可用于移动的和物联网场景的快速和高性能解决方案。该计划将为相关学生提供坚实的理论基础和实用的编程技能，使他们为未来的图像/视频处理和计算机视觉职业做好准备。他们将为加拿大经济做出宝贵贡献，帮助加拿大在人工智能及其应用领域保持世界领先地位。