Research map / Image generation / Image restoration
Unveiling Details in the Dark: Simultaneous Brightening and Zooming for Low-Light Image Enhancement
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 BrZoNet, a Retinex-based siamese decoupling network plus an illumination-aware interaction unit and multi-stream SR cage that jointly brightens and super-resolves low-light images, trained and evaluated on RELLISUR.
Problem
Super-resolution of low-light, low-resolution images: existing SR methods trained on normal-light data fail to brighten, produce noise, artifacts and color bias, and cascading enhancement with SR propagates errors.
Approach
Decompose low-light and normal-light images into reflectance and illuminance maps with a shared-weight siamese Retinex decoupling network, then fuse the two branches via a gated cross-layer interaction unit (IaIU) and a multi-stream SR cage, combining results by element-wise multiplication.
Key results
- On RELLISUR ×2, BrZoNet reaches 22.79 PSNR / 0.745 SSIM / 0.243 LPIPS, ranking first on all six metrics versus HAT (20.21 PSNR at ×2 for the second-best ranked entry in table order is not the strongest baseline; best baseline PSNR is 21.21 from SwinIR).
- On RELLISUR ×4, BrZoNet reaches 21.41 PSNR, versus 20.29 for the best baseline (SwinIR).
- Ablation shows IllNet, IaIU and MSC each improve ×2 PSNR (21.91 to 22.79 combined).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add IllNet to the model without IllNet, IaIU, or MSC (ablation) | BrZoNet · RELLISUR ×2 | PSNR: 21.91 -> 22.16 | Helped |
| Add IaIU to IllNet (ablation) | BrZoNet · RELLISUR ×2 | PSNR: 22.16 -> 22.31 | Helped |
| Add MSC to IllNet + IaIU (ablation) | BrZoNet · RELLISUR ×2 | PSNR: 22.31 -> 22.79 | Helped |
| Add IllNet to the model without IllNet, IaIU, or MSC (ablation) | BrZoNet · RELLISUR ×4 | PSNR: 20.78 -> 20.90 | Helped |
| Add IaIU to IllNet (ablation) | BrZoNet · RELLISUR ×4 | PSNR: 20.90 -> 21.16 | Helped |
| Add MSC to IllNet + IaIU (ablation) | BrZoNet · RELLISUR ×4 | PSNR: 21.16 -> 21.42 | Helped |
| Use siamese decoupling that also decomposes ynhr (ablation) | BrZoNet · RELLISUR ×2 | PSNR: 22.46 -> 22.79 | Helped |
| Use siamese decoupling that also decomposes ynhr (ablation) | BrZoNet · RELLISUR ×4 | PSNR: 21.25 -> 21.42 | Helped |
| Remove perceptual loss (ablation) | BrZoNet · RELLISUR | detail recovery | Hurt |
| Remove Unet skip connections (ablation) | BrZoNet · RELLISUR ×2 | PSNR: 22.79 -> 22.17 | Hurt |
| Change loss weights to 0.5/0.5/0.5 (ablation) | BrZoNet · RELLISUR ×2 | PSNR: 22.79 -> 22.33 | Hurt |
| Change loss weights to 1/0.1/0.1 (ablation) | BrZoNet · RELLISUR ×2 | PSNR: 22.79 -> 22.57 | Hurt |
Surprising findings
- Cascading a low-light enhancer (SCI or LLFormer) with HAT yields noisy results with insufficient detail, so cascade and direct SR are both sub-optimal.
Limitations
- Evaluated only on a single dataset (RELLISUR) at ×2 and ×4 scales.
- Relies on paired normal-light high-resolution ground truth for siamese decoupling supervision.
Experimental setup
Datasets: RELLISUR
Benchmarks: RELLISUR
Metrics: PSNR SSIM LPIPS RMSE FSIM SRE
Baselines: HAT SRFormer SwinIR Restormer MIRNet EDSR RCAN ESRGAN SRFBN SCI+HAT LLFormer+HAT
Builds on: Retinex theory HAT MIRNet SKFF
Closest papers by meaning
- Luminance-Aware Statistical Quantization: Unsupervised Hierarchical Learning for Illumination Enhancement (NeurIPS 2025 · similarity 0.86)
- VQCNIR: Clearer Night Image Restoration with Vector-Quantized Codebook (AAAI 2024 · similarity 0.85)
- Interpretable Unsupervised Joint Denoising and Enhancement for Real-World low-light Scenarios (ICLR 2025 · similarity 0.85)
- Low-Light Face Super-resolution via Illumination, Structure, and Texture Associated Representation (AAAI 2024 · similarity 0.85)
- Towards Robust Event-guided Low-Light Image Enhancement: A Large-Scale Real-World Event-Image Dataset and Novel Approach (CVPR 2024 · similarity 0.85)
- See through the Dark: Learning Illumination-affined Representations for Nighttime Occupancy Prediction (NeurIPS 2025 · similarity 0.84)
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.