基于深度学习的智能手机实时视频超分辨率：Mobile AI 2021 挑战赛报告
图像与视频处理
2021-05-20 v1 计算机视觉与模式识别
机器学习
摘要
随着视频通信与流媒体服务的兴起，视频超分辨率近来已成为最重要的移动端相关问题之一。尽管已提出许多针对该任务的解决方案，但其中大多数计算开销过大，无法在硬件资源有限的便携设备上运行。为解决此问题，我们引入了首个 Mobile AI 挑战赛，其目标是开发基于端到端深度学习的视频超分辨率解决方案，能在移动端 GPU 上实现实时性能。参与者获提供 REDS 数据集，并训练其模型以进行高效的 4 倍视频放大。所有模型的运行时间在搭载骁龙 865 SoC 的 OPPO Find X2 智能手机上评估，该 SoC 可借助其 Adreno GPU 加速浮点网络。所提出的解决方案完全兼容任意移动端 GPU，并能在高达 80 FPS 下将视频放大至高清分辨率，同时展现高保真结果。本文提供了挑战赛中开发的所有模型的详尽描述。
引用
@article{arxiv.2105.08826,
title = {Real-Time Video Super-Resolution on Smartphones with Deep Learning, Mobile AI 2021 Challenge: Report},
author = {Andrey Ignatov and Andres Romero and Heewon Kim and Radu Timofte and Chiu Man Ho and Zibo Meng and Kyoung Mu Lee and Yuxiang Chen and Yutong Wang and Zeyu Long and Chenhao Wang and Yifei Chen and Boshen Xu and Shuhang Gu and Lixin Duan and Wen Li and Wang Bofei and Zhang Diankai and Zheng Chengjian and Liu Shaoli and Gao Si and Zhang Xiaofeng and Lu Kaidi and Xu Tianyu and Zheng Hui and Xinbo Gao and Xiumei Wang and Jiaming Guo and Xueyi Zhou and Hao Jia and Youliang Yan},
journal= {arXiv preprint arXiv:2105.08826},
year = {2021}
}
备注
Mobile AI 2021 Workshop and Challenges: https://ai-benchmark.com/workshops/mai/2021/. arXiv admin note: substantial text overlap with arXiv:2105.07825. substantial text overlap with arXiv:2105.08629, arXiv:2105.07809, arXiv:2105.08630