Research map / Label-efficient and robust learning / Self-supervised and contrastive
Disentangle Nighttime Lens Flares: Self-supervised Generation-based Lens Flare Removal
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes a physics-based nighttime lens flare formation model and SGLFR-Net, a self-supervised, training-data-free network that jointly removes glow and reflective (ghost) lens flares.
Problem
Removing co-occurring lens flares (glow plus reflective/ghost flares) from nighttime photos, which existing single-type or cascaded methods fail to handle because the flares share a light source and are formed jointly.
Approach
A nighttime lens flare formation model linking glow (scattering PSF) and reflective flare (optical symmetry) from the same light source, realized by a two-stream self-supervised network: PSF-rendering network for glow and a texture-prior (inpainting-style) network for reflections, constrained by an optical-symmetry texture prior module, with no pre-training or training data.
Key results
- On OurSynDatasets for joint glow and ghost removal, the method reaches 29.65 PSNR and 0.974 SSIM, the best in the table, without any training data.
- On BracketFlare reflection removal, the self-supervised method achieves SSIM 0.967 and LPIPS 0.038, second only to the supervised CVPR2023 method (0.994 / 0.004) in SSIM and LPIPS.
- Ablation: removing the PSFR-Net stream drops PSNR from 29.65 to 27.78 on the joint task.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use SGLFR-Net instead of the strongest reported glow-and-reflection pipeline | joint glow and reflective flare removal · OurSynDatasets | PSNR: 29.23 -> 29.65 | Helped |
| Use SGLFR-Net instead of the strongest reported glow-and-reflection pipeline | joint glow and reflective flare removal · OurSynDatasets | SSIM: 0.970 -> 0.974 | Helped |
| Use SGLFR-Net instead of Dai (CVPR2023) | reflective flare removal · BracketFlare | SSIM: 0.994 -> 0.967 | Hurt |
| Use SGLFR-Net instead of Dai (CVPR2023) | reflective flare removal · BracketFlare | LPIPS: 0.004 -> 0.038 | Hurt |
| Add PSFR-Net (ablation) | SGLFR-Net · OurSynDatasets joint flare removal | PSNR: 27.78 -> 29.65 | Helped |
| Add PSFR-Net (ablation) | SGLFR-Net · OurSynDatasets joint flare removal | SSIM: 0.973 -> 0.974 | Helped |
| Add TPRR-Net (ablation) | SGLFR-Net · OurSynDatasets joint flare removal | PSNR: 29.61 -> 29.65 | Helped |
| Add TPRR-Net (ablation) | SGLFR-Net · OurSynDatasets joint flare removal | SSIM: 0.972 -> 0.974 | Helped |
| Add 1−L_SSIM to the loss (ablation) | SGLFR-Net · OurSynDatasets joint flare removal | PSNR: 28.18 -> 29.65 | Helped |
| Add 1−L_SSIM to the loss (ablation) | SGLFR-Net · OurSynDatasets joint flare removal | SSIM: 0.965 -> 0.974 | Helped |
| Add L_MSE to the loss (ablation) | SGLFR-Net · OurSynDatasets joint flare removal | PSNR: 29.37 -> 29.65 | Helped |
| Add L_MSE to the loss (ablation) | SGLFR-Net · OurSynDatasets joint flare removal | SSIM: 0.972 -> 0.974 | Helped |
Surprising findings
- The self-supervised method, with no training data, is competitive with fully supervised reflection-removal baselines on BracketFlare, though it does not beat the best supervised model.
Limitations
- Joint evaluation relies on a self-synthesized dataset (OurSynDatasets) since no joint flare dataset exists, and real-world joint results lack ground truth.
- Relies on hand-designed priors (PSF rendering, optical symmetry) whose assumptions may not hold for all lens configurations.
- Per-image optimization approach; lower performance than the best supervised method on some metrics.
Experimental setup
Datasets: BracketFlare OurSynDatasets (synthesized joint glow and ghost flare data from BracketFlare images) ECCV2022 light effect dataset real-world captured nighttime flare images
Benchmarks: OurSynDatasets BracketFlare ECCV2022 light effect dataset real-world captured images
Metrics: PSNR SSIM LPIPS
Baselines: Flare7K++ ICCV2023 glow suppression ACM MM2023 ECCV2022 light effect suppression CVPR2024 CVPR2018 reflection removal CVPR2020 reflection removal ICCV2021 reflection removal CVPR2023 (Dai et al.)
Builds on: Flare7K Deep Image Prior Criminisi exemplar-based inpainting APSF
Closest papers by meaning
- FlareX: A Physics-Informed Dataset for Lens Flare Removal via 2D Synthesis and 3D Rendering (NeurIPS 2025 · similarity 0.92)
- Language-Guided One-Step Diffusion Model for Nighttime Flare Removal (CVPR 2026 · similarity 0.88)
- OmniSR: Shadow Removal Under Direct and Indirect Lighting (AAAI 2025 · similarity 0.84)
- Recasting Regional Lighting for Shadow Removal (AAAI 2024 · similarity 0.83)
- FLARE: Feed-forward Geometry, Appearance and Camera Estimation from Uncalibrated Sparse Views (CVPR 2025 · similarity 0.82)
- Flash-Split: 2D Reflection Removal with Flash Cues and Latent Diffusion Separation (CVPR 2025 · similarity 0.81)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.