Papers › Improving Image Restoration by Revisiting Global Information Aggregation

Improving Image Restoration by Revisiting Global Information Aggregation

8 Dec 2021arXiv:2112.04491archive 2025-07-28

Xiaojie Chu, Liangyu Chen, Chengpeng Chen, Xin Lu

Global operations, such as global average pooling, are widely used in top-performance image restorers. They aggregate global information from input features along entire spatial dimensions but behave differently during training and inference in image restoration tasks: they are based on different regions, namely the cropped patches (from images) and the full-resolution images. This paper revisits global information aggregation and finds that the image-based features during inference have a different distribution than the patch-based features during training. This train-test inconsistency negatively impacts the performance of models, which is severely overlooked by previous works. To reduce the inconsistency and improve test-time performance, we propose a simple method called Test-time Local Converter (TLC). Our TLC converts global operations to local ones only during inference so that they aggregate features within local spatial regions rather than the entire large images. The proposed method can be applied to various global modules (e.g., normalization, channel and spatial attention) with negligible costs. Without the need for any fine-tuning, TLC improves state-of-the-art results on several image restoration tasks, including single-image motion deblurring, video deblurring, defocus deblurring, and image denoising. In particular, with TLC, our Restormer-Local improves the state-of-the-art result in single image deblurring from 32.92 dB to 33.57 dB on GoPro dataset. The code is available at https://github.com/megvii-research/tlc.

PaperPDFCode

In Syntology Open this paper in Syntology's Atlas, the map of the papers in Syntology's graph and their citations.

Code

megvii-research/TLC officialmentioned in papermentioned on GitHubpytorchNOASSERTION report
megvii-research/NAFNet mentioned on GitHubpytorch report
setsunil/dsdnet mentioned on GitHubpytorchNOASSERTION report

Repository list and official/mentioned flags are the archive's, frozen 2025-07-28. Reachability, where shown, is from one Syntology probe window (2026-09-16 to 2026-09-18); repositories not probed show nothing. GitHub stars are not tracked.

Code Syntology ran Syntology

Not run by Syntology. Nothing on this page verifies that the listed code works.

Tasks

Color Image DenoisingDeblurringDenoisingGrayscale Image DenoisingImage DeblurringImage Defocus DeblurringImage DehazingImage DenoisingImage RestorationSemantic SegmentationSingle Image DeblurringVideo Deblurring

Results from the paper archive 2025-07-28

TaskDatasetModelMetricValueRank at snapshotLeaderboardReport
Color Image Denoising Urban100 sigma30 Restormer-Local PSNR 33.06 #1 of 2 Archive leaderboard report
Color Image Denoising Urban100 sigma50 Restormer-Local PSNR 30.17 #3 of 9 Archive leaderboard report
Deblurring GoPro RNN-MBP-Local PSNR 33.8 #13 of 56 Archive leaderboard report
Deblurring GoPro RNN-MBP-Local SSIM 0.966 #13 of 56 Archive leaderboard report
Deblurring GoPro Restormer-Local PSNR 33.57 #17 of 56 Archive leaderboard report
Deblurring GoPro Restormer-Local SSIM 0.966 #17 of 56 Archive leaderboard report
Deblurring GoPro MPRNet-local PSNR 33.31 #22 of 56 Archive leaderboard report
Deblurring GoPro MPRNet-local SSIM 0.964 #22 of 56 Archive leaderboard report
Deblurring GoPro HINet-local PSNR 33.08 #24 of 56 Archive leaderboard report
Deblurring GoPro HINet-local SSIM 0.962 #24 of 56 Archive leaderboard report
Deblurring HIDE (trained on GOPRO) Restormer-TLC PSNR (sRGB) 31.49 #12 of 26 Archive leaderboard report
Deblurring HIDE (trained on GOPRO) Restormer-TLC Params (M) 26.13 #12 of 26 Archive leaderboard report
Deblurring HIDE (trained on GOPRO) Restormer-TLC SSIM (sRGB) 0.945 #12 of 26 Archive leaderboard report
Deblurring HIDE (trained on GOPRO) MPRNet-TLC PSNR (sRGB) 31.19 #15 of 26 Archive leaderboard report
Deblurring HIDE (trained on GOPRO) MPRNet-TLC Params (M) 20.1 #15 of 26 Archive leaderboard report
Deblurring HIDE (trained on GOPRO) MPRNet-TLC SSIM (sRGB) 0.942 #15 of 26 Archive leaderboard report
Deblurring MSU BASED MPR local ERQAv2.0 0.74521 #8 of 11 Archive leaderboard report
Deblurring MSU BASED MPR local LPIPS 0.08323 #8 of 11 Archive leaderboard report
Deblurring MSU BASED MPR local PSNR 31.65037 #8 of 11 Archive leaderboard report
Deblurring MSU BASED MPR local SSIM 0.94542 #8 of 11 Archive leaderboard report
Deblurring MSU BASED MPR local Subjective 0.4407 #8 of 11 Archive leaderboard report
Deblurring MSU BASED MPR local VMAF 67.01788 #8 of 11 Archive leaderboard report
Grayscale Image Denoising Urban100 sigma25 Restormer-Local PSNR 31.55 #2 of 10 Archive leaderboard report
Grayscale Image Denoising Urban100 sigma50 Restormer-Local PSNR 28.41 #2 of 10 Archive leaderboard report
Grayscale Image Denoising urban100 sigma15 Restormer-Local PSNR 33.85 #2 of 3 Archive leaderboard report
Image Deblurring GoPro Restormer-TLC PSNR 33.57 #20 of 55 Archive leaderboard report
Image Deblurring GoPro Restormer-TLC Params (M) 26.13 #20 of 55 Archive leaderboard report
Image Deblurring GoPro Restormer-TLC SSIM 0.966 #20 of 55 Archive leaderboard report
Image Deblurring GoPro MPRNet-TLC PSNR 33.31 #22 of 55 Archive leaderboard report
Image Deblurring GoPro MPRNet-TLC Params (M) 20.1 #22 of 55 Archive leaderboard report
Image Deblurring GoPro MPRNet-TLC SSIM 0.964 #22 of 55 Archive leaderboard report
Image Deblurring GoPro HINet-TLC PSNR 33.08 #28 of 55 Archive leaderboard report
Image Deblurring GoPro HINet-TLC SSIM 0.962 #28 of 55 Archive leaderboard report

Ranks are positions in the archive's leaderboards as they stood at the 2025-07-28 snapshot. Results published since then are not among these rows, so a rank here is not a current standing.

Methods

Introduced by this paper: TLC

Average PoolingGlobal Average PoolingInstance NormalizationTLC

Report a problem or propose a change · a person checks every report against the paper or source before anything changes; decisions are listed on /corrections