Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Evaluating Low-Light Image Enhancement Across Multiple Intensity Levels

Maria Pilligua, David Serrano-Lozano, Pai Peng, Ramon Baldrich, Michael S. Brown, Javier Vazquez-Corral

TL;DR

The paper tackles the limited understanding of LLIE performance across different illumination levels by introducing the MILL dataset, which captures each scene at 11 illumination levels with fixed camera parameters and lux-measured illuminance, totaling 1100 RAW images. It analyzes a wide range of state-of-the-art LLIE methods under varying brightness and shows large performance fluctuations across intensity ranges. To boost robustness, it introduces two auxiliary losses—Intensity Prediction Loss and Scene Content Loss—that encourage disentanglement of illumination from scene content and are integrated with a Retinexformer baseline, yielding substantial improvements (up to $10$ dB PSNR on DSLR Full HD and $2$ dB on smartphone). Overall, MILL provides a rigorous cross-illumination benchmark and the proposed losses offer a principled way to train LLIE models that generalize across diverse lighting.

Abstract

Imaging in low-light environments is challenging due to reduced scene radiance, which leads to elevated sensor noise and reduced color saturation. Most learning-based low-light enhancement methods rely on paired training data captured under a single low-light condition and a well-lit reference. The lack of radiance diversity limits our understanding of how enhancement techniques perform across varying illumination intensities. We introduce the Multi-Illumination Low-Light (MILL) dataset, containing images captured at diverse light intensities under controlled conditions with fixed camera settings and precise illuminance measurements. MILL enables comprehensive evaluation of enhancement algorithms across variable lighting conditions. We benchmark several state-of-the-art methods and reveal significant performance variations across intensity levels. Leveraging the unique multi-illumination structure of our dataset, we propose improvements that enhance robustness across diverse illumination scenarios. Our modifications achieve up to 10 dB PSNR improvement for DSLR and 2 dB for the smartphone on Full HD images.

Evaluating Low-Light Image Enhancement Across Multiple Intensity Levels

TL;DR

The paper tackles the limited understanding of LLIE performance across different illumination levels by introducing the MILL dataset, which captures each scene at 11 illumination levels with fixed camera parameters and lux-measured illuminance, totaling 1100 RAW images. It analyzes a wide range of state-of-the-art LLIE methods under varying brightness and shows large performance fluctuations across intensity ranges. To boost robustness, it introduces two auxiliary losses—Intensity Prediction Loss and Scene Content Loss—that encourage disentanglement of illumination from scene content and are integrated with a Retinexformer baseline, yielding substantial improvements (up to dB PSNR on DSLR Full HD and dB on smartphone). Overall, MILL provides a rigorous cross-illumination benchmark and the proposed losses offer a principled way to train LLIE models that generalize across diverse lighting.

Abstract

Imaging in low-light environments is challenging due to reduced scene radiance, which leads to elevated sensor noise and reduced color saturation. Most learning-based low-light enhancement methods rely on paired training data captured under a single low-light condition and a well-lit reference. The lack of radiance diversity limits our understanding of how enhancement techniques perform across varying illumination intensities. We introduce the Multi-Illumination Low-Light (MILL) dataset, containing images captured at diverse light intensities under controlled conditions with fixed camera settings and precise illuminance measurements. MILL enables comprehensive evaluation of enhancement algorithms across variable lighting conditions. We benchmark several state-of-the-art methods and reveal significant performance variations across intensity levels. Leveraging the unique multi-illumination structure of our dataset, we propose improvements that enhance robustness across diverse illumination scenarios. Our modifications achieve up to 10 dB PSNR improvement for DSLR and 2 dB for the smartphone on Full HD images.

Paper Structure

This paper contains 15 sections, 3 equations, 5 figures, 4 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Low-Light Datasets
  4. Low-Light Image Enhancement Methods
  5. Multi-Illumination Low-Light (MILL) Dataset
  6. Method using New Loss Terms
  7. Intensity Prediction Loss
  8. Scene Content Loss
  9. Combined Objective Function
  10. Experiments
  11. Benchmark on MILL
  12. Qualitative Results
  13. FullHD Experiments and Ablation
  14. Outdoor Images
  15. Conclusion

Figures (5)

  • Figure 1: Impact of brightness variation on LLIE model performance. Blending input images with ground truth at 20% and 50% ratios degrades Retinexformer performance.
  • Figure 2: Example scenes from our dataset at different levels for both the DSLR camera (first and second rows) and the smartphone camera (third row). We show the three lowest values to illustrate how intensity varies across successive levels. We also display higher levels to show they remain noticeably underexposed compared to the ground truth.
  • Figure 3: Visual comparison on MILL-s. From left to right: input, SCI SCI, GT-Mean liao2025gt, PromptNorm promptnorm, Retinexformer Retinexformer, Ours, and ground truth. We show three examples with zoomed-in regions below each. First two images: DSLR; last image: smartphone.
  • Figure 4: Ablation Study for the different components of our loss term. On the second row, we show the $\Delta E_{76}$ error maps.
  • Figure 5: Outdoor examples from the DICM DICM (first row) and SICE SICE (second row) of Retinexformer trained on LoLv1, our baseline with the two proposed additional loss terms independently, and our final approach.