One deblurring network for every camera: Fourier convolution learning lifts GoPro PSNR to 34.02 dB

Motion deblurring now runs through one network across different cameras and scenes, instead of a specialized model trained per benchmark dataset. Existing deblurring networks mostly specialize on isolated benchmark datasets and drop sharply when deployed on unseen cameras or content; All-in-One restoration models mix denoising, deraining and deblurring in one network, but denoising and deraining are additive corruptions while motion blur is a convolutional degradation, so joint optimization causes conflicts and feature interference. Under a unified training protocol, UniDeblur-L raises GoPro PSNR from EVSSM's 33.26 dB to 34.02 dB, HIDE from 31.70 dB to 32.49 dB, and RSBlur from 33.91 dB to 34.37 dB.

The network stops relying only on pixel-to-pixel mapping and instead learns an input-adaptive global filter in Fourier space, replacing spatial kernel estimation with element-wise multiplication. Mainstream networks use spatial pixel-level black-box mapping that fits image content rather than the convolution process; earlier frequency-domain methods either treat frequency representations as additive residual features or learn a static global filter that is degradation-agnostic. In ablations, attaching the Fourier Convolution Learning module at encoder-decoder transitions raises GoPro PSNR from the 31.96 dB baseline to above 32.26 dB with ConvBlock, MLPBlock, GatedBlock or SS2DBlock, at roughly 1.32G extra FLOPs and about 1GB extra memory.

Training and evaluation move to multi-source data with a pruning and quality-control pipeline that removes spatially misaligned pairs and ground truths with residual blur. Prior methods lean on the synthetic GoPro dataset, which holds only 3.2K blur-sharp pairs across 33 scenes and just 2.1K training pairs, lacking real-world semantic and kernel variation and inviting overfitting. The training corpus combines GoPro, HIDE, REDS, RSBlur, RealBlur and LOLBlur, with BSD1/BSD2/BSD3, MIORe, VAR-MIORe and FoundIR as out-of-distribution tests; on BSD2, UniDeblur-B reaches CLIPIQA 0.3031 and MUSIQ 51.17 against EVSSM's 0.2856 and 47.88.

The report covers unified motion deblurring: one network trained on six in-domain datasets and evaluated on six out-of-distribution test sets, with PSNR, SSIM and the no-reference CLIPIQA and MUSIQ metrics. It suits researchers and engineers working on image restoration, frequency-domain methods and cross-domain generalization. Code is planned for release at github.com/DeepMed-Lab-ECNU/Single-Image-Deblur, along with test datasets and experimental results.

Cross-domain evaluation leans mainly on no-reference metrics such as CLIPIQA and MUSIQ, because real-world ground-truth quality degrades with exposure time, with ground-truth MUSIQ on BSD falling from 55.05 to 49.70. Ablations show that pushing the Fourier Convolution Learning module's depth from 4 to 8 yields only small gains while memory rises to 46GB. ReLoBlur and RealLOL appear only as diagnostic references and stay out of the main comparison, because the former shows color shifts and residual blur in ground truths and the latter lacks corresponding ground truths and is dominated by extremely dark scenes.