NTIRE 2020 Challenge on Real Image Denoising: Dataset, Methods and Results

Abdelhamed, Abdelrahman; Afifi, Mahmoud; Timofte, Radu; Brown, Michael S.; Cao, Yong; Zhang, Zhilu; Zuo, Wangmeng; Liu, Jiye; Chen, Wendong; Wen, Changyuan; Li, Meng; Lv, Shuailin; Zhang, Yunchao; Pan, Zhihong; Li, Baopu; Xi, Teng; Fan, Yanwen; Yu, Xiyu; Zhang, Gang; Liu, Jingtuo; Han, Junyu; Ding, Errui; Yu, Songhyun; Park, Bumjun; Liu, Shuai; Zong, Ziyao; Nan, Nan; Li, Chenghua; Yang, Zhiping; Bao, Lei; Wang, Shuangquan; Bai, Dingqun; Lee, Jung Won; Kim, Young-Jung; Rho, Kyeongha; Shin, Chanho; Kim, Sungho; Tang, Pengliang; Zhao, Yi; Zhou, Yuqian; Yi, Fan; Huang, Thomas S.; Li, Zhihao; Shah, Nisarg; Liu, Wei; Yan, Qing; Zhao, Yuzhi; Możejko, Marcin; Latkowski, Tomasz; Treszczotko, Lukasz; Michał, Szafraniuk,; Trojanowski, Krzysztof; Wu, Yue; Michelini, Pablo Navarrete; Hu, Fengshuo; Lu, Yunhua; Kim, Sujin; Kim, Won-Jin; Lee, Jaayeon; Choi, Jang-Hwan; Zhussip, Magauiya; Khassenov, Azamat; Kim, Jong Hyun; Cho, Hwechul; Kansal, Priya; Nathan, Sabari; Ye, Zhangyu; Lu, Xiwen; Wu, Yiyan; Yang, Jiangxin; Tang, Siliang; Cao, Yanan; Maggioni, Matteo; Marras, Ioannis; Tanay, Thomas; Slabaugh, Greg; Yan, Yunfei; Kang, Myungjoo; Choi, Hyoung Gwon; Song, Kyung Ja; Xu, Shusong; Lu, Xiaomu; Wang, Tingniao; Lei, Chunxia; Liu, Bin; Gupta, R.; Kumar, V.

doi:10.48550/arxiv.2005.04117

Cited by 1 publication

(3 citation statements)

References 0 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…Performance on real-world denoising As stated from the beginning, ES-WMV is designed with real-world IPs in mind, targeting unknown noise types and levels. Given the encouraging performance above, we test it on the standard RGB track of NTIRE 2020 Real Image Denoising Challenge [1]. Just to be clear, our purpose here is not to compete with SOTA methods on this dataset-they are datadriven methods based on the training set provided in the challenge, or to prove the supremacy of DIP for denoising.…”

Section: Image Denoisingmentioning

confidence: 99%

“…parameter. (1) Here, the loss is chosen according to the noise model, and the regularizer R encodes priors on x. The advent of deep learning (DL) has revolutionized how IPs are solved: on the radical side, deep neural networks (DNNs) are trained to directly map any given y to an x; on the mild side, pretrained DL models are taken to replace certain nonlinear mappings in optimization algorithms to solve (1) (e.g., plugand-play, and algorithm unrolling).…”

Section: Introductionmentioning

confidence: 99%

“…The smallest PSNR gaps within each cell are in red. 1. 1.02/15.4 1.13 (0.47) 15.6 4.94 25.2/16.8 0.35 (0.08) 10.9 7.80 1.54/21.0 0.56 (0.62) 11.8 7.72 0.39/12.5 0.87 (0.14) Peppers 2.91 6.07 3.69/8.12 0.60 (0.37) 5.27 2.65 7.27/13.2 0.36 (0.34) 5.28 6.84 5.47/9.53 0.35 (0.24) 3.89 6.69 1.76/7.64 0.98 (0.43) Lena 5.68 7.79 3.08/10.3 0.27 (0.11) 8.04 8.36 8.73/17.2 0.39 (0.32) 8.62 6.54 0.73/8.94 0.50 (0.47) 6.17 7.67 0.92/8.62 0.07 (0.06) Baboon 7.73 0.33 2.17/5.78 1.36 (0.12) 3.90 6.27 8.18/0.53 0.27 (0.09) 7.48 1.79 1.74/0.81 1.63 (0.10) 6.66 3.25 1.73/0.35 1.07 (0.26) F16 5.17 9.10 3.75/0.89 0.44 (0.07) 9.29 1.57 7.38/2.27 0.12 (0.21) 5.65 6.39 3.59/2.13 0.31 (0.13) 3.32 5.75 1.65/0.50 0.34 (0.17) Kodak1 8.65 2.84 15.0/3.17 1.92 (0.15) 11.0 0.22 8.32/1.74 0.82 (0.40) 0.64 6.43 1.75/3.63 2.28 (0.78) 7.45 6.08 11.0/2.83 1.81 (0.11) Kodak2 0.66 6.11 0.64/7.55 2.51 (0.58) 13.6 0.97 7.95/13.4 2.00 (0.28) 9.55 11.9 0.69/10.9 0.63 (0.07) 11.0 7.06 1.02/10.6 2.46 (0.71)…”

mentioning

confidence: 99%

See 2 more Smart Citations

Early Stopping for Deep Image Prior

Wang¹,

Li²,

Zhang³

et al. 2021

Preprint

View full text Add to dashboard Cite

Deep image prior (DIP) and its variants have showed remarkable potential for solving inverse problems in computer vision, without any extra training data. Practical DIP models are often substantially overparameterized. During the fitting process, these models learn mostly the desired visual content first, and then pick up the potential modeling and observational noise, i.e., overfitting. Thus, the practicality of DIP often depends critically on good early stopping (ES) that captures the transition period. In this regard, the majority of DIP works for vision tasks only demonstrates the potential of the models-reporting the peak performance against the ground truth, but provides no clue about how to operationally obtain near-peak performance without access to the groundtruth. In this paper, we set to break this practicality barrier of DIP, and propose an efficient ES strategy, which consistently detects near-peak performance across several vision tasks and DIP variants. Based on a simple measure of dispersion of consecutive DIP reconstructions, our ES method not only outpaces the existing ones-which only work in very narrow domains, but also remains effective when combined with a number of methods that try to mitigate the overfitting. The code is available at https://github.com/sunumn/Early_Stopping_for_DIP.

show abstract

Section: Image Denoisingmentioning

confidence: 99%

Section: Introductionmentioning

confidence: 99%

mentioning

confidence: 99%

See 1 more Smart Citation