NTIRE 2020 Challenge on Real-World Image Super-Resolution: Methods and Results

1. Introduction

• Super-resolution increases image resolution by adding high-frequency content, with recent deep-learning work emphasizing perceptual appeal alongside earlier PSNR-focused methods.The paper notes that deep CNN methods address the ill-posed task, with adversarial losses among techniques used for perceptual results.Source: §1 · p. 1
• Paired low- and high-resolution training images are generally unavailable in real-world applications, while bicubic downscaling changes image characteristics such as sensor noise and harms generalization.The challenge instead learns from unpaired data without restricting assumptions about input image formation.Source: §1 · pp. 1–2
• The challenge provides source images and unpaired high-quality targets, with human-study rankings for perceptual quality in both tracks.Track 1 uses an unknown synthetic degradation and supports reference-based metrics; Track 2 uses low-quality smartphone images without ground truth.Source: §1 · p. 2
• Track 1 uses inputs degraded by unknown image-processing operations, whereas Track 2 uses low-quality smartphone-camera images without ground truth.The synthetic degradation in Track 1 permits reference-based evaluation; both tracks target perceptually pleasing results.Source: §1 · p. 2

2. NTIRE 2020 Challenge

• The challenge sought to promote weak and unsupervised super-resolution research, establish a benchmark protocol and dataset, and probe the field’s state of the art; it was organized on Codalab.Both tracks target 4× upscaling.Source: §2 · p. 2

2.1. Track 1: Image Processing Artifacts

• Track 1 uses artificial degradation to enable reference-based evaluation, with a learned, undisclosed operator intended to generalize across degradations and input domains.Training inputs are degraded Flickr2K images without downsampling, paired only with clean DIV2K target images; validation and test inputs are downscaled before degradation, with original DIV2K images as ground truth.Source: §2.1 · p. 2

2.2. Track 2: Smartphone Images

• Track 2 super-resolves real low-quality smartphone images, using iPhone3 images from DPED as source data and Track 1’s target images.Because no ground truth exists above sensor size, crops from DPED’s validation set are used for a human perception study.Source: §2.2 · p. 2

2.3. Challenge phases

• The challenge had development, validation, and final-test phases, progressing from training and validation images to final submissions of results, descriptions, code, and model weights.During validation, participants could submit Track 1 results for PSNR and SSIM measurement, and consult a validation leaderboard.Source: §2.3 · pp. 2–3

3. Challenge Results

• Track 1 drew 19 valid methods from 16 teams among 292 registrants, while Track 2 drew 15 valid methods from 14 teams among 251 registrants.Teams submitted results, code or executables, and factsheets before the final test phase ended.Source: §3 · p. 3

3.1. Architectures and Main Ideas

• Most top methods learned a degradation operator to generate paired training data before training super-resolution, and results suggest training strategy and degradation modeling mattered more than architecture choice.Popular architectures included RRDB/ESRGAN and RCAN, while many methods used adversarial and perceptual VGG losses.Source: §3.1 · p. 3
• Notable variations included AITA-Noah’s alternating degradation/SR training, Impressionism’s blur-kernel and noise estimation, and Samsung-SLSI-MSL’s robust training across sampled degradations.AITA-Noah also used explicit denoising and a sharpening network; Impressionism used KernelGAN for Track 2 kernel estimation, which improved sharpness and quality in generated SR images.Source: §3.1 · p. 3

3.2. Baselines

• Track 1 benchmarks included Bicubic, pretrained RRDB, and supervised ESRGAN as an upper-bound comparison; Track 2 had no supervised baseline because ground-truth HR images were unavailable.The ESRGAN baseline was fine-tuned using the synthetic Track 1 degradation, while pretrained RRDB used clean images and bicubic downsampling.Source: §3.2 · p. 3

3.3. Evaluation Metrics

• Final rankings relied on Amazon Mechanical Turk: Track 1 used reference-based MOS, while Track 2 used MOR because ground-truth references were unavailable.Track 1 also reported PSNR, SSIM, and LPIPS, but LPIPS was treated only as an indication of perceptual quality; Track 2 reported NIQE, BRISQUE, PIQE, NRQM, PI, and average IQA-Rank. Because IQA metrics did not correlate well with perceived Track 2 quality, organizers visually selected the five methods for the MOR study.Source: §3.3 · pp. 3–5

3.4. Track 1: Image Processing Artifacts

• Impressionism led Track 1 with MOS 9.5% better than Samsung-SLSI-MSL; all top-10 methods beat Bicubic, and Impressionism surpassed supervised ESRGAN.ITS425 had the second-best PSNR and best SSIM; third-ranked BOE-IOT-AIBD also achieved very impressive PSNR and SSIM.Source: Table 1 · p. 4, §3.4 · p. 5

3.5. Track 2: Smartphone Images

• Impressionism outperformed other Track 2 approaches by a large margin in MOR and achieved the best average IQA-Rank.Its images were sharper, with almost no noise and few artifacts; AITA-Noah and ITS425 produced clean but less sharp images, and kernel estimation was credited for Impressionism’s detail. Relbmag-Eht had a similar IQA-Rank despite structured noise; Samsung-SLSI-MSL was unranked after using paired DSLR images and still lacked comparable sharpness. Impressionism’s output also had better image quality than the camera source.Source: §3.5 · pp. 5–7, Table 2 · p. 5

4.1. Impressionism; 4.2. AITA-Noah

• Impressionism synthesizes paired training data using estimated blur kernels and collected real-noise patches, then trains an ESRGAN-based model with pixel, perceptual, and adversarial losses.AITA-Noah uses iterative degradation learning for Track 1, but was unranked because it trained on source validation and test images; its Track 2 CycleSR approach combines image translation and supervised SR to reduce color and brightness changes.Source: §4.1 · p. 8, §4.2 · pp. 8–9

4.9. KU-ISPL; 4.10. Webbzhou; 4.11. SR-DL

• KU-ISPL uses an unpaired three-generator framework, Webbzhou separates frequency bands in degradation learning, and SR-DL jointly denoises and super-resolves with dSRVAE.KU-ISPL assigns noise, color, and texture losses to separate discriminators. Webbzhou uses a Color Attention Residual Block to mitigate color shift and an EdgeLoss with a Canny operator; SR-DL adds an SRSN to a denoising autoencoder and uses a discriminator to align SR outputs with reference-image appearance.Source: §4.9 · p. 11, §4.10 · p. 12, §4.11 · p. 12

4.12. TeamAY; 4.13. Bigfeature-Camera; 4.14. BMIPL-UNIST-YH

• TeamAY’s SimUSR creates pseudo-pairs from multiple LR images, Bigfeature-Camera trains multi-scale RNAN on DSGAN-generated data, and BMIPL-UNIST-YH pairs CycleGAN degradation learning with RCAN.SimUSR uses pretrained RCAN and BM3D, enabling supervised-learning techniques without runtime training; Bigfeature adds contrast-channel attention and local-attention blocks to RNAN. BMIPL follows a two-stage unpaired-data approach.Source: §4.12 · p. 12, §4.13 · pp. 12–13, §4.14 · p. 13

4.15. SVNIT1; 4.16. SVNIT2; 4.17. KU-ISPL2

• SVNIT1 combines self-supervised pixel loss with unsupervised quality losses, SVNIT2 enforces consistency between SR and downscaling networks, and KU-ISPL2 extends SRGAN with multiscale convolutions.SVNIT1 submitted adversarial-loss and variational-encoder variants; SVNIT2 also uses VGG, GAN, and learned image-quality losses.Source: §4.15 · p. 13, §4.16 · p. 13, §4.17 · p. 13

4.18. SuperT; 4.19. GDUT-wp; 4.20. MLP-SR

• SuperT uses a balanced Laplacian pyramid, GDUT-wp ensembles bicubic-trained SRResNets, and MLP-SR combines DSGAN training with VGG, GAN, TV, and L1 losses.GDUT-wp selects from its ensemble to reduce random artifacts, while the MLP-SR method also uses test-time ensembling.Source: §4.18 · p. 13, §4.19 · p. 13, §4.20 · p. 13

4.21. Relbmag-Eht; 4.22. QCAM

• Relbmag-Eht integrates degradation matching into an ESRGAN model with LR- and HR-phase discriminators, while QCAM fine-tunes a pretrained SR model using low-resolution supervision.QCAM minimizes the mismatch between bicubic-downsampled SR output and the source image.Source: §4.21 · p. 13, §4.22 · p. 13

• Training used unpaired source images with unknown degradations and clean, high-quality target images, rather than paired high- and low-resolution images.Source: §5 · p. 13
• The challenge covered Track 1 image-processing artifacts and Track 2 low-quality smartphone images, with 22 teams competing in the final step.Source: §5 · p. 13
• The challenge aimed to stimulate unsupervised super-resolution research by establishing a standard benchmark and new baseline methods.The paper notes that most participants were influenced by AIM 2019 and demonstrated innovative solutions.Source: §5 · p. 13

ITS425

• ITS425’s submission was titled “Addptive Residual dense block network for real image super-resolution.”Source: ITS425 · p. 15