Research map / Image generation / Image restoration
Thermal-Aware Low-Light Image Enhancement: A Real-World Benchmark and a New Light-Weight Model
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.
Introduces TLIE, a real-world dataset of 1,113 low-light/normal-light/thermal triplets, and ThermNet, a light-weight multimodal network that fuses thermal maps into low-light image enhancement via SNR alignment, multimodal Retinex decomposition, and thermal-guided refinement.
Problem
Low-light image enhancement from RGB alone struggles in very dark regions, and there is no real-world public dataset pairing low-light RGB, normal-light images, and thermal maps for multimodal LIE.
Approach
Fuse thermal maps into LIE via three modules: a Multimodal SNR Alignment Module that fuses low-light and thermal SNR maps, a multimodal Thermal-based Retinex Network that decomposes both inputs into reflectance and illumination, and a Thermal-guided Refinement Network with a Multi-scale Thermal Fusion Layer in each decoder stage.
Key results
- ThermNet reaches 19.81 PSNR, 0.468 SSIM and 0.556 LPIPS on TLIE with 0.51M parameters, beating Retinexformer (19.42 PSNR) and ELIE (19.50 PSNR, 220M parameters).
- Ablation: MTFL alone raises PSNR from 18.53 to 19.48; the full model with all three modules reaches 19.81.
- Plugging MTFL into prior methods improves most of them, e.g. DeepLPF PSNR 17.63 to 19.74 and LEDNet 19.33 to 19.55 in the reported table.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use ThermNet instead of Retinexformer | Low-light image enhancement on TLIE · TLIE | PSNR: 19.42 -> 19.81 | Helped |
| Add MTRN (ablation) | ThermNet · TLIE test set | PSNR: 18.53 -> 19.46 | Helped |
| Add MSAM (ablation) | ThermNet · TLIE test set | PSNR: 18.53 -> 19.26 | Helped |
| Add MTFL (ablation) | ThermNet · TLIE test set | PSNR: 18.53 -> 19.48 | Helped |
| Add MTRN and MSAM (ablation) | ThermNet · TLIE test set | PSNR: 18.53 -> 19.63 | Helped |
| Add MTRN and MTFL (ablation) | ThermNet · TLIE test set | PSNR: 18.53 -> 19.74 | Helped |
| Add MSAM and MTFL (ablation) | ThermNet · TLIE test set | PSNR: 18.53 -> 19.55 | Helped |
| Add MTRN, MSAM, and MTFL (ablation) | ThermNet · TLIE test set | PSNR: 18.53 -> 19.81 | Helped |
Surprising findings
- Adding MTFL to DeepLPF yields a large PSNR gain (about 2.1 dB) and to DNF a smaller one, while the paper reports that the gain is not uniform across methods, with some degraded metrics.
Limitations
- Dataset is small (1,113 samples) and collected with a single thermal camera (Hikvision K20) and iPhone rig.
- Low-light images are synthesized by adjusting exposure and ISO on normal scenes rather than captured naturally dark scenes.
- Training uses only MSE-based losses and evaluation is limited to TLIE with no cross-dataset tests.
Experimental setup
Datasets: TLIE
Benchmarks: TLIE
Metrics: PSNR SSIM MSSM LPIPS
Baselines: EFINet ChebyLighter DeepLPF DNF MIRNet SNR-Net LEDNet Restormer Retinexformer eSL-Net ELIE
Builds on: Retinex-Net SNR-Net Restormer Retinexformer
Closest papers by meaning
- Enhancing Infrared Vision: Progressive Prompt Fusion Network and Benchmark (NeurIPS 2025 · similarity 0.85)
- Enhancing Low-Light Images: A Synthetic Data Perspective on Practical and Generalizable Solutions (AAAI 2025 · similarity 0.84)
- ThermalGen: Style-Disentangled Flow-Based Generative Models for RGB-to-Thermal Image Translation (NeurIPS 2025 · similarity 0.84)
- Zero-Shot Low-Light Image Enhancement via Latent Diffusion Models (AAAI 2025 · similarity 0.84)
- Luminance-Aware Statistical Quantization: Unsupervised Hierarchical Learning for Illumination Enhancement (NeurIPS 2025 · similarity 0.83)
- AGLLDiff: Guiding Diffusion Models Towards Unsupervised Training-free Real-world Low-light Image Enhancement (AAAI 2025 · similarity 0.83)
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.