第九届NTIRE 2024高效超分辨率挑战赛报告
计算机视觉与模式识别
2024-06-26 v2 图像与视频处理
摘要
本文全面回顾了NTIRE 2024挑战赛，重点关注高效单图像超分辨率(ESR)解决方案及其成果。该挑战赛的任务是基于低分辨率图像和对应的高分辨率图像对，以放大因子x4对输入图像进行超分辨率处理。主要目标是开发在运行时间、参数量和FLOPs等各方面进行优化的网络，同时保持DIV2K_LSDIR_valid数据集上峰值信噪比(PSNR)约为26.90 dB，DIV2K_LSDIR_test数据集上约为26.99 dB。此外，本次挑战赛有4个赛道，包括主赛道（整体性能）、子赛道1（运行时间）、子赛道2（FLOPs）和子赛道3（参数量）。在主赛道中，考虑了所有三个指标（即运行时间、FLOPs和参数量）。主赛道的排名基于所有其他子赛道得分的加权和计算。在子赛道1中，评估了提交方案的实际运行时间性能，并使用相应得分确定排名。在子赛道2中，考虑了FLOPs数量。基于相应FLOPs计算的得分用于确定排名。在子赛道3中，考虑了参数量。基于相应参数量计算的得分用于确定排名。RLFN被设置为效率测量的基线。挑战赛有262名注册参与者，34支队伍提交了有效方案。他们评估了高效单图像超分辨率的最新技术。为促进挑战赛的可重复性并使其他研究人员能够基于这些发现进行构建，验证方案的代码和预训练模型已在https://github.com/Amazingren/NTIRE2024_ESR/公开。
引用
@article{arxiv.2404.10343,
title = {The Ninth NTIRE 2024 Efficient Super-Resolution Challenge Report},
author = {Bin Ren and Yawei Li and Nancy Mehta and Radu Timofte and Hongyuan Yu and Cheng Wan and Yuxin Hong and Bingnan Han and Zhuoyuan Wu and Yajun Zou and Yuqing Liu and Jizhe Li and Keji He and Chao Fan and Heng Zhang and Xiaolin Zhang and Xuanwu Yin and Kunlong Zuo and Bohao Liao and Peizhe Xia and Long Peng and Zhibo Du and Xin Di and Wangkai Li and Yang Wang and Wei Zhai and Renjing Pei and Jiaming Guo and Songcen Xu and Yang Cao and Zhengjun Zha and Yan Wang and Yi Liu and Qing Wang and Gang Zhang and Liou Zhang and Shijie Zhao and Long Sun and Jinshan Pan and Jiangxin Dong and Jinhui Tang and Xin Liu and Min Yan and Qian Wang and Menghan Zhou and Yiqiang Yan and Yixuan Liu and Wensong Chan and Dehua Tang and Dong Zhou and Li Wang and Lu Tian and Barsoum Emad and Bohan Jia and Junbo Qiao and Yunshuai Zhou and Yun Zhang and Wei Li and Shaohui Lin and Shenglong Zhou and Binbin Chen and Jincheng Liao and Suiyi Zhao and Zhao Zhang and Bo Wang and Yan Luo and Yanyan Wei and Feng Li and Mingshen Wang and Yawei Li and Jinhan Guan and Dehua Hu and Jiawei Yu and Qisheng Xu and Tao Sun and Long Lan and Kele Xu and Xin Lin and Jingtong Yue and Lehan Yang and Shiyi Du and Lu Qi and Chao Ren and Zeyu Han and Yuhan Wang and Chaolin Chen and Haobo Li and Mingjun Zheng and Zhongbao Yang and Lianhong Song and Xingzhuo Yan and Minghan Fu and Jingyi Zhang and Baiang Li and Qi Zhu and Xiaogang Xu and Dan Guo and Chunle Guo and Jiadi Chen and Huanhuan Long and Chunjiang Duanmu and Xiaoyan Lei and Jie Liu and Weilin Jia and Weifeng Cao and Wenlong Zhang and Yanyu Mao and Ruilong Guo and Nihao Zhang and Qian Wang and Manoj Pandey and Maksym Chernozhukov and Giang Le and Shuli Cheng and Hongyuan Wang and Ziyan Wei and Qingting Tang and Liejun Wang and Yongming Li and Yanhui Guo and Hao Xu and Akram Khatami-Rizi and Ahmad Mahmoudi-Aznaveh and Chih-Chung Hsu and Chia-Ming Lee and Yi-Shiuan Chou and Amogh Joshi and Nikhil Akalwadi and Sampada Malagi and Palani Yashaswini and Chaitra Desai and Ramesh Ashok Tabib and Ujwala Patil and Uma Mudenagudi},
journal= {arXiv preprint arXiv:2404.10343},
year = {2024}
}
备注
The report paper of NTIRE2024 Efficient Super-resolution, accepted by CVPRW2024