PaperlayerView exact revision on arXiv

Source-linked AI summary

LensBridge: Frequency-Guided Compound Degradation Adaptation for Lens Aberration Correction and Veiling Glare Removal

Xiaolong Qian, Zhonghua Yi, Qi Jiang, Kailun Yang, Shuhang Xie, Shaohua Gao, Kaiwei Wang

arXiv:2610.00318v1eess.IVcs.CVphysics.optics

TL;DR

Simplified optical systems can combine lens aberrations with veiling glare, while reusable LensLib foundations are trained for aberrations and realistic compound training pairs are difficult to obtain. LensBridge builds a PSF-aware correction foundation and adapts it using frequency-guided data and model procedures; experiments report favorable restoration performance across optical systems.

  • Problem

    Reusable LensLib foundations cover aberrations but not veiling glare, and realistic target-system compound pairs are difficult to obtain.

    Source: §1 · pp. 1–2

  • Method

    LensBridge builds a PSF-aware LensLib aberration foundation, then uses shared frequency guidance for FDC compound-pair synthesis and FPD pseudo-observation adaptation with few unpaired target images.

    Source: §1 · p. 2

  • Results

    Across multiple optical systems, experiments report favorable restoration performance, effective few-shot adaptation, and applicability across restoration backbones.

    Source: §5 · p. 10

  • Takeaways & Limitations

    The approach adapts reusable aberration correction to joint aberration and veiling-glare removal without target-system paired supervision.

    Source: §1 · p. 3

    LensBridge relies on LensLib coverage, and its VG adaptation uses frequency-guided completion rather than a fully calibrated physical model.

    Source: §5 · p. 10

Abstract

from arXiv · show

Simplified optical systems often exhibit residual lens aberrations and Veiling Glare (VG), resulting in spatially varying blur and contrast reduction. Large-scale Lens Libraries (LensLib) enable reusable aberration correction models by covering diverse Point Spread Functions (PSFs), but their aberration-only training distribution does not include target-specific veiling glare. Extending such foundations to compound degradation is challenging because realistic target-system compound pairs are difficult to obtain. To address this challenge, we propose LensBridge, a two-stage framework that first establishes a reusable aberration correction foundation and then adapts it to compound optical degradation using only a few unpaired target observations. In Stage I, we build a PSF-aware one-step diffusion foundation by constructing discrete degradation priors from LensLib PSFs and learning to retrieve them directly from aberrated images, enabling PSF-aware correction without requiring explicit PSF at inference. In Stage II, we adapt this foundation to compound degradation through frequency-domain guidance. At the data level, Frequency-guided Degradation Completion (FDC) transfers target low-frequency characteristics to LensLib aberrated images while preserving aberration structures to synthesize compound training pairs; at the model level, Frequency-guided Pseudo Decomposition (FPD) forms aberration- and VG-dominant pseudo observations to condition separate adaptation branches. Extensive experiments across multiple optical systems demonstrate that LensBridge effectively extends reusable aberration correction foundations to joint aberration correction and veiling glare removal without target-system paired supervision. All code will be available at https://github.com/XiaolongQian/LensBridge.

1 INTRODUCTION

LensBridge adapts reusable aberration correction to compound aberration and veiling-glare degradation using a few unpaired target observations. It combines PSF-aware priors with frequency-guided data and model adaptation, addressing a gap in aberration-only foundations that are difficult to extend without realistic compound training pairs.

1 INTRODUCTION

  • LensLib-based foundations cover diverse aberrations but not veiling glare, while realistic target-system compound pairs are difficult to obtain.LensLib pretraining supports reusable correction models across aberration distributions; the paper pursues foundation-to-compound adaptation rather than target-specific synthesis.Source: §1 · p. 2
  • Aberration changes image amplitude across a broad frequency range, whereas veiling glare primarily changes low-frequency amplitude, motivating frequency-guided adaptation.The illustrated example also reports a smaller phase-swap reconstruction error for the additional veiling glare.Source: §1 · p. 2, Figure 1 · p. 2
  • Stage I encodes LensLib PSFs as discrete degradation priors and retrieves them from aberrated observations to condition a one-step diffusion restoration backbone.The retrieved priors provide PSF-aware guidance without requiring an explicit PSF at inference.Source: §1 · p. 2
  • FDC transfers target low-frequency degradation characteristics to LensLib aberrated images while preserving aberration structures and clean counterparts to construct compound training pairs.It uses a shared target-system Fourier-amplitude statistic and does not require an additional degradation generator.Source: §1 · p. 2
  • FPD uses the shared frequency statistic to form aberration-dominant and VG-dominant pseudo observations that separately condition two adaptation branches.Together, FDC supplies target-like training supervision and FPD separately conditions the aberration and VG branches.Source: §1 · p. 3
  • LensBridge extends a reusable LensLib aberration foundation to compound degradation using few unpaired target observations, with experiments across optical systems demonstrating effectiveness.The framework introduces FDC for target-like compound pairs and FPD for degradation-dominant pseudo observations.Source: §1 · p. 3

2 RELATED WORK

Related work spans lens-aberration correction, veiling-glare removal, and frequency-guided adaptation. LensBridge distinguishes its approach by reusing a LensLib aberration foundation while adapting to compound degradation with unpaired target observations.

Lens Aberration Correction

  • Learning-based CAC methods use simulated optical degradations but often generalize poorly to unseen systems; LensLib pretraining improves cross-lens reuse, yet existing frameworks address PSF-induced aberrations rather than veiling glare.The related work also describes earlier deconvolution and PSF-based inverse filtering approaches.Source: §2 · p. 3

Veiling Glare Removal

  • Veiling-glare methods include physical modeling, image decomposition, and data-driven synthesis; DeVeiler synthesizes target-specific compound degradations but repeats that modeling for each optical system.Veiling glare is spatially distributed and dominated by low-frequency variations, unlike the localized structures typical of lens flare.Source: §2 · p. 3

Frequency-Guided Degradation Adaptation

  • Prior frequency-guided methods use statistics for global domain alignment or single-degradation adaptation, whereas LensBridge completes VG characteristics while preserving LensLib aberration structure and reuses the statistic for separate pseudo observations.The overview depicts a reusable aberration foundation followed by compound-degradation adaptation with separate branches for aberration- and VG-dominant pseudo observations.Source: §2 · p. 3, Figure 2 · p. 4

3 METHODOLOGY

LensBridge formulates compound optical restoration as adapting a reusable LensLib aberration foundation using few unpaired target observations. It combines PSF-derived prior retrieval with frequency-guided data completion and separate aberration- and glare-conditioned adaptation branches.

3.1 PROBLEM FORMULATION

  • The degradation model represents each color-channel patch with a local PSF, contrast attenuation, and additive glare.It assumes the PSF is shift-invariant within each spatial patch.Source: §3.1 · p. 4
  • The method first learns a reusable aberration foundation from LensLib source pairs, then adapts it to compound target degradation using only unpaired target images.The target set is small relative to the source dataset, and target-system compound pairs are not required.Source: §3.1 · p. 4, §3.2 · p. 4

3.3 STAGE I: REUSABLE ABERRATION CORRECTION FOUNDATION

  • LensBridge uses SD-Turbo as a one-step diffusion backbone, conditioning its LoRA adaptation on discrete aberration priors retrieved from aberrated images.The priors are constructed from LensLib PSFs; an image-based predictor retrieves them without a PSF at inference.Source: §3.2 · p. 4, §3.3 · p. 5

3.4 STAGE II: TARGET-SYSTEM COMPOUND DEGRADATION ADAPTATION

  • A shared target Fourier-amplitude statistic guides completion of LensLib aberrated images and separate adaptation branches conditioned on aberration- and VG-dominant pseudo observations.FDC transfers target characteristics while retaining aberration structure; the pseudo observations provide factor-specific cues rather than physically exact degradation components.Source: §3.4 · pp. 5–6

4 EXPERIMENTS

Across unpaired and paired target settings, LensBridge performs favorably on the evaluated optical systems, while ablations support its frequency-guided data and model adaptation components. Few-shot results and tests with other systems and backbones indicate broader applicability within the reported experiments.

4.1 EXPERIMENT SETTINGS

  • The evaluated systems include SL and MRL, with Screen and Real domains, plus CAYE, Nano, and UDC as additional real-world cases.The primary adaptation setting uses unpaired compound-degraded target images; a paired setting is also evaluated for comparison.Source: §4.1 · pp. 6–7, §4.2 · p. 7

4.2 COMPARISON WITH STATE-OF-THE-ART METHODS

  • In the unpaired-target setting, LensBridge performs favorably on Screen and Real domains, with consistent gains in LPIPS and no-reference metrics, including over DeVeiler.Visual comparisons also show recovered fine details and improved color contrast on both domains.Source: Table 1 · p. 7, §4.2 · p. 8
  • On CAYE and nano-optics, Stage I recovers aberration-degraded details, while target adaptation further suppresses veiling glare and restores color and contrast.Each system is adapted using only a few unpaired degraded observations.Source: §4.2 · p. 8, Figure 5 · p. 8
  • In the target aberration-paired setting, LensBridge achieves the best overall performance across both optical systems and domains.The comparison includes cascaded pipelines and restoration and adaptation methods.Source: §4.2 · pp. 8–9, Table 2 · p. 8

4.3 ABLATION STUDIES

  • FDC achieves the best performance among tested synthesis strategies using aggregated target-domain spectral statistics, and consistently improves SwinIR and S3Diff.The strategy comparison uses Screen-SL; backbone results are reported on Screen SL.Source: §4.3 · p. 9, Table 3 · p. 9, Figure 6 · p. 9
  • Component ablations show that FPD-guided separate degradation conditioning achieves the best overall performance; adding VG LoRA without FPD yields limited improvement under entangled degradations.FDC-based adaptation substantially improves the baseline, while longer backbone training brings limited gains.Source: §4.3 · p. 9, Table 4 · p. 9
  • On Screen SL, a few unpaired target images yield substantial gains over the non-adapted model, while additional images bring diminishing improvements.The study varies the number of unpaired target images.Source: §4.3 · p. 9, Figure 7 · p. 9
  • On Real MRL, aggregating more target observations stabilizes the spectral statistic used by FDC and FPD to construct compound-like inputs and degradation-dominant pseudo observations.The reported analysis supports the statistic as a target cue for compound-degradation adaptation.Source: §4.3 · p. 9, Figure 8 · p. 9

5 CONCLUSION

LensBridge adapts a reusable aberration-correction foundation to compound aberration and veiling-glare degradation using a shared Fourier-amplitude statistic to guide FDC and FPD. The authors report favorable restoration, few-shot adaptation, and applicability across multiple systems and backbones, while noting limits in foundation coverage and physical VG modeling.

5 CONCLUSION

  • LensBridge adapts a reusable aberration-correction foundation to compound degradation using a shared Fourier-amplitude statistic to guide FDC and FPD.The paper reports experiments across multiple optical systems, few-shot adaptation, and different restoration backbones.Source: §5 · p. 10
  • The foundation's coverage limits unseen-lens adaptation, and VG adaptation uses frequency-guided completion rather than a fully calibrated physical model.The authors identify broader target-specific aberration priors and physically grounded VG modeling under complex illumination as future directions.Source: §5 · p. 10

A ADDITIONAL IMPLEMENTATION DETAILS

Implementation uses two 30K-iteration phases to establish the Stage-I prior space and train image-based prior retrieval, then adapts each target system independently for 3K iterations. FDC-generated compound pairs are sampled alongside LensLib aberration-only pairs during Stage-II adaptation.

A ADDITIONAL IMPLEMENTATION DETAILS

  • Stage I has two 30K-iteration phases: the first learns the PSF-derived discrete prior space, and the second freezes its encoder and codebook while training image-based retrieval with teacher supervision.The Phase-1 objective is Lrec + LVQ; Phase 2 uses Lrec + LFM.Source: A ADDITIONAL IMPLEMENTATION DETAILS · p. 15
  • Stage II initializes from Stage I and independently adapts each optical system for 3K iterations, jointly optimizing the image-based predictor, conditioning modules, VG encoder, and both LoRA branches.The pretrained SD-Turbo weights and discrete codebook remain frozen.Source: A ADDITIONAL IMPLEMENTATION DETAILS · p. 15
  • Adaptation jointly samples FDC-generated compound pairs and original LensLib aberration-only pairs; feature matching is retained because FPD's aberration-dominant observations are pseudo observations.Tab. 5 summarizes unpaired target-observation counts across systems, including SL and MRL in Screen and Real domains and additional real-world cases.Source: A ADDITIONAL IMPLEMENTATION DETAILS · p. 15, Table 5 · p. 15

B ADDITIONAL ANALYSIS

Blind prior retrieval preserves lens-aware degradation information without PSFs at inference, while aggregated target spectral statistics and spatial modulation support stable, diverse degradation completion.

Blind Prior Retrieval

  • Image-retrieved priors cluster around their PSF-derived anchors and remain separated across optical systems, indicating lens-aware retrieval without PSFs at inference.The visualization covers held-out single-element, smartphone, and meta-optical lenses.Source: B ADDITIONAL ANALYSIS · p. 15, Figure 9 · p. 16

Analysis of Target Spectral Statistics

  • Aggregating target observations reduces scene-specific amplitude variation; radial profiles converge and their distance to the 50-shot statistic decreases as more images are included.Single-image amplitude transfer introduces content-dependent structures into recomposed images.Source: Figure 10 · p. 16, B ADDITIONAL ANALYSIS · p. 16

Frequency-Domain Interpretation of Veiling Glare

  • Veiling glare produces stronger low-frequency amplitude variations, whereas the aberration optical transfer function affects image structures across frequency ranges.This spectral distinction motivates target-domain amplitude statistics for FDC and FPD.Source: B ADDITIONAL ANALYSIS · p. 16

Effect of Spatial Modulation

  • Spatial modulation maintains comparable reconstruction fidelity while improving perceptual quality, particularly MUSIQ and Q-Align, and supports more diverse synthesized compound degradation.The comparison is on Screen-SL.Source: B ADDITIONAL ANALYSIS · p. 16, Table 6 · p. 17

C.1 ADDITIONAL QUANTITATIVE RESULTS

Additional experiments compare LensBridge with one-step diffusion restorers and examine few-shot adaptation and sensitivity to the clear reference dataset across specified domains.

Complete Results under the Target Aberration-Paired Setting

  • The complete target aberration-paired quantitative results cover Screen and Real domains and compare cascaded, dehazing, and aberration-adaptation methods.The compared methods include SwinIR cascaded with Flare7K++ or DiffDehaze, DiffDehaze, QDMR, DeVeiler, and OmniLens.Source: §C.1 · p. 17

Comparison with One-Step Diffusion Restoration Models

  • LensBridge achieves consistently favorable performance, particularly on perceptual and no-reference image quality metrics, versus OSEDiff and S3Diff across four SL/MRL domains and two adaptation settings.The settings are unpaired-target and target aberration-paired.Source: §C.1 · p. 17, Table 8 · p. 18

Target Data Efficiency

  • With five unpaired target observations, PSNR rises from 15.075 to 19.551 and LPIPS falls from 0.332 to 0.277 on Screen-SL.Further increases in target-image count yield smaller incremental improvements.Source: §C.1 · pp. 17–18, Table 9 · p. 18

Sensitivity to the Clear Reference Dataset

  • On Real-SL, results remain comparable when DIV2K is replaced with Flickr2K as the clear reference collection, indicating limited sensitivity to that choice.Table 10 evaluates sensitivity to the clear reference dataset.Source: §C.1 · p. 18

C.2 ADDITIONAL VISUAL COMPARISONS

Visual comparisons cover unpaired and aberration-paired SL/MRL settings, additional optical systems, real UDC images, and failure cases.

Visual Comparisons under the Unpaired-Target Setting

  • Across Screen and Real domains and both SL and MRL systems, LensBridge recovers fine structures while reducing contrast and color degradation under the unpaired-target setting.The reported improvements span controlled screen captures and in-the-wild observations.Source: §C.2 · p. 18, Figure 11 · p. 19, Figure 12 · p. 20

Results on Additional Optical Systems

  • On CAYE and nano-optics, Stage I alleviates aberration blur and recovers details, while target adaptation further improves contrast and color under compound degradation.Each system is adapted independently using a few unpaired degraded observations.Source: §C.2 · pp. 18–19, Figure 13 · p. 20

Visual Comparisons under the Target Aberration-Paired Setting

  • Under target aberration-paired supervision, LensBridge produces sharper structures and improved contrast than existing restoration and adaptation methods across SL and MRL systems.The visual comparisons complement quantitative results under the same setting.Source: §C.2 · p. 19, Figure 14 · p. 21, Figure 15 · p. 21

Additional Results on Real UDC Images

  • On real UDC images, LensBridge improves global contrast and visibility, but fine diffraction artifacts remain challenging without diffraction-aware PSFs in LensLib.The evaluation is a qualitative case beyond the current LensLib degradation distribution.Source: §C.2 · p. 19, Figure 16 · p. 22

Failure Cases and Future Work

  • Fine high-frequency text remains difficult to reconstruct in SL, while severe MRL overexposure causes information loss that prevents faithful recovery.The paper attributes the SL difficulty partly to latent-space compression and the lack of dedicated high-frequency restoration designs.Source: §C.2 · p. 19, Figure 17 · p. 22

Related papers

Loading 2610.00318v1…