Statistics

AI Image Restoration Statistics: Datasets, Benchmarks, and Model Results

Key statistics on AI image restoration datasets, deblurring, denoising, document repair, quality metrics, and processing speed.

AI image restoration spans deblurring, denoising, document repair, and related tasks that turn damaged or degraded pixels into more usable images. The available benchmark results show both scale and trade-offs: LSDIR contains 84,991 training images, while NTIRE deblurring entries range from sub-second processing to 464.8 seconds per image. Scores also depend heavily on the task and metric.

Contents

Dataset scale

LSDIR is one of the clearest indicators of how much training data can be assembled for image restoration. Its dataset contains 84,991 training images, 1,000 validation images, and 1,000 test images. The LSDIR source describes the 84,991-image training set as high quality and contrasts it with earlier restoration datasets containing only some thousands of images. These figures describe dataset composition, not the number of images an individual restoration tool can process.

The split is deliberately uneven. Training uses 84,991 images, while validation and testing each use 1,000. That gives the training portion a much larger role in model fitting, while the two smaller partitions provide fixed-sized checks for development and evaluation. The source label for these figures is LSDIR.

Dataset size is not the same as task coverage. A large collection can support learning broad visual patterns, but benchmark scores still depend on the degradation being repaired, the evaluation images, and the metric being reported. The results below therefore keep deblurring, denoising, and document restoration separate.

NTIRE deblurring benchmarks

The NTIRE 2021 Challenge on Image Deblurring reported results for two tracks. Track 1’s leading entry, VIDAR, scored 29.04 PSNR, 0.8416 SSIM, and 0.2397 LPIPS, with a reported runtime of 1.0 second per image. The runner-up, netai, scored 28.91 PSNR, 0.8246 SSIM, and 0.2569 LPIPS in 12.4 seconds. Third-place NJUST-IMAG recorded 28.51 PSNR, 0.8172 SSIM, and 0.2547 LPIPS in 6.4 seconds.

The remaining reported Track 1 entries were:

TeamPSNRSSIMLPIPSRuntime
SRC-B28.440.81580.25310.9 s
Baidu28.440.81350.270440.8 s
MMM28.420.81320.268514.3 s
Imagination28.360.81300.26667.3 s
Noah CVlab28.330.81320.260624.5 s
TeamInception28.280.81100.26510.9 s
ZOCS Team28.250.81080.26362.2 s
Mier28.210.81090.264617.3 s
INFINITY28.110.80640.27342.7 s
DMLAB27.870.80090.28302.3 s
RTQSA-Lab27.780.79600.28306.5 s
Yonsei-MCML27.640.79560.27301.6 s
SCUT-ZS27.610.79360.28850.3 s

The Track 1 bicubic upsampling baseline scored 24.06 PSNR, 0.6817 SSIM, and 0.5120 LPIPS; no runtime was reported for it. In this benchmark, the baseline’s PSNR and SSIM were below every listed team result, while its LPIPS was higher. Those are benchmark comparisons, not a universal ranking for every image or restoration workflow.

Speed and quality trade-offs

Track 2 of the NTIRE 2021 deblurring challenge shows a different balance between quality and processing time. The winner, The Fat, The Thin and The Strong, scored 29.70 PSNR, 0.8403 SSIM, and 0.2319 LPIPS, but reported 464.8 seconds per image. Noah CVlab scored 29.62 PSNR, 0.8397 SSIM, and 0.2304 LPIPS in 76.1 seconds. CAPP OB scored 29.60 PSNR, 0.8398 SSIM, and 0.2302 LPIPS in 12.7 seconds.

Track 2 teamPSNRSSIMLPIPSRuntime
Baidu29.590.83810.234071.0 s
SRC-B29.560.83850.23220.8 s
Mier29.340.83550.254617.3 s
VIDAR29.330.85650.22225.3 s
DuLang29.170.83250.2411not reported
TeamInception29.110.82920.244910.1 s
GiantPandaCV29.070.82860.24992.4 s
Maradona28.960.82640.250621.4 s
LAB FUD28.920.82590.2424not reported
SYJ28.810.82220.25461.4 s
Dseny28.260.80810.26030.6 s
IPCV IITM27.910.80280.29476.4 s
DMLAB27.840.80130.293433.2 s
Blur Attack27.410.78870.31241.7 s

The Track 2 no-processing baseline scored 24.94 PSNR, 0.7199 SSIM, and 0.3265 LPIPS, with no runtime reported. VIDAR’s Track 2 result is notable because its SSIM of 0.8565 was higher than the winner’s 0.8403, while its LPIPS of 0.2222 was lower than the winner’s 0.2319 and its runtime was 5.3 seconds. Different metrics can therefore identify different strengths within the same benchmark.

RAW image denoising

The MIPI 2024 Challenge on Few-shot RAW Image Denoising measured a different restoration problem. It had 165 registered participants and 7 final teams. MiVideoNR won with 31.225 PSNR, 0.9542 SSIM, and a challenge score of 30.9645. Samsung placed second with 30.965 PSNR, 0.9545 SSIM, and 30.7064 score.

The remaining final teams were AIIA, with 29.981 PSNR, 0.9487 SSIM, and 29.6886 score; MS-Denoimer, with 29.926 PSNR, 0.9481 SSIM, and 29.6300 score; MyTurn, with 29.926 PSNR, 0.9447 SSIM, and 29.6096 score; HBNU, with 29.734 PSNR, 0.9462 SSIM, and 29.4263 score; and Erlong Mountain Team, with 23.433 PSNR, 0.9121 SSIM, and 22.9211 score.

These figures should not be directly ranked against the NTIRE deblurring values as though they measured the same task. RAW denoising, image deblurring, and document repair have different inputs, distortions, evaluation sets, and challenge scoring conventions. The MIPI source label is MIPI 2024 Challenge on Few-shot RAW Image Denoising.

Document restoration data

DocRes reports a generalist document image restoration setup across multiple tasks. Its unified training ran for 100,000 steps on 8 NVIDIA A6000 GPUs, with a global batch size of 80 and a maximum learning rate of 2e-4. Before unified training, DocRes was pre-trained for 50,000 steps on dewarping.

Each of the five tasks had a sampling weight of 0.2 during unified training. Four tasks used 256x256 random crops, while dewarping used 256x256 resizing. These settings describe the training procedure rather than a consumer application’s speed or output resolution.

The dewarping data included 100K synthetic Doc3D samples and 300 real DIR300 distorted images. Deshadowing used 14,200 synthetic FSDSRD images and 4,371 real RDD images. Its evaluation sets included Jung’s 87-image set, Kligler’s 300-image set, and OSR’s 237-image set. Appearance enhancement used 90K synthetic Doc3DShade images and 450 real RealDAE images, with 150 RealDAE test images and 130 DocUNet images for evaluation.

For text dewarping, DocRes used the 66K-sample TDD dataset, selected 40K training samples, and evaluated on 1.6K testing samples. For binarization, it used DIBCO’18 as the test set and trained on the remaining H-DIBCO years plus Noisy Office, Synchromedia Multispectral, Persian Heritage Image Binarization, and Bickley Diary datasets. The source label for these training and dataset figures is DocRes: A Generalist Model Toward Unifying Document Image Restoration Tasks.

Training and evaluation statistics

DocRes’s reported dewarping result on DIR300 was 0.6264 MS-SSIM, 0.241 AD, and 6.83 LD. On Kligler et al.’s deshadowing set, it scored 0.9005 SSIM and 27.14 PSNR. Because the evaluation sets and metrics differ, these values are best read as task-specific measurements rather than a single overall restoration score.

Across the supplied benchmark results, three patterns stand out. First, training-data scale varies widely, from hundreds of real distorted document images in DIR300 to tens of thousands of samples in LSDIR, Doc3D, TDD, and related synthetic datasets. Second, quality metrics move independently: a result can have a stronger SSIM while another has a lower LPIPS or faster runtime. Third, runtime is a practical dimension of restoration. NTIRE Track 2 includes entries from 0.6 seconds to 464.8 seconds per image, so the highest score is not automatically the most convenient choice for interactive photo editing.

For readers evaluating AI photo restoration tools, the most useful comparison is therefore task-specific. Look for deblurring results when the problem is camera shake or motion blur, denoising results for sensor noise, and document-restoration results for shadows, warping, or binarization. Keep the measurement period and source label attached to each number: the NTIRE figures are from its 2021 deblurring challenge, the denoising figures are from MIPI’s 2024 challenge, and the DocRes figures describe its reported training and evaluation experiments.

Written by

unblurai.com Editorial Team

Editorial team

unblurai.com publishes practical how-to guides and educational articles with clear steps and useful context.