Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

TPCNet: Triple physical constraints for Low-light Image Enhancement

Jing-Yi Shi, Ming-Fei Li, Ling-An Wu

TL;DR

Low-light image enhancement is improved by moving beyond Lambertian Retinex constraints to Kubelka-Munk based triple physical constraints (TPC) that connect illumination, specular reflectance, and detection. These constraints are embedded in an implicit feature-space network (TPCNet) featuring Light/Reflectivity Feature Estimators, a Color-Association Mechanism, and a Dual-Stream Cross-Guided Transformer for efficient, long-range feature fusion. The approach yields state-of-the-art results with modest parameter count and FLOPs across ten datasets, while maintaining color fidelity and robustness to varying color spaces. This work offers a principled, interpretable route to robust LLIE with strong generalization and computational efficiency.

Abstract

Low-light image enhancement is an essential computer vision task to improve image contrast and to decrease the effects of color bias and noise. Many existing interpretable deep-learning algorithms exploit the Retinex theory as the basis of model design. However, previous Retinex-based algorithms, that consider reflected objects as ideal Lambertian ignore specular reflection in the modeling process and construct the physical constraints in image space, limiting generalization of the model. To address this issue, we preserve the specular reflection coefficient and reformulate the original physical constraints in the imaging process based on the Kubelka-Munk theory, thereby constructing constraint relationship between illumination, reflection, and detection, the so-called triple physical constraints (TPCs)theory. Based on this theory, the physical constraints are constructed in the feature space of the model to obtain the TPC network (TPCNet). Comprehensive quantitative and qualitative benchmark and ablation experiments confirm that these constraints effectively improve the performance metrics and visual quality without introducing new parameters, and demonstrate that our TPCNet outperforms other state-of-the-art methods on 10 datasets.

TPCNet: Triple physical constraints for Low-light Image Enhancement

TL;DR

Low-light image enhancement is improved by moving beyond Lambertian Retinex constraints to Kubelka-Munk based triple physical constraints (TPC) that connect illumination, specular reflectance, and detection. These constraints are embedded in an implicit feature-space network (TPCNet) featuring Light/Reflectivity Feature Estimators, a Color-Association Mechanism, and a Dual-Stream Cross-Guided Transformer for efficient, long-range feature fusion. The approach yields state-of-the-art results with modest parameter count and FLOPs across ten datasets, while maintaining color fidelity and robustness to varying color spaces. This work offers a principled, interpretable route to robust LLIE with strong generalization and computational efficiency.

Abstract

Low-light image enhancement is an essential computer vision task to improve image contrast and to decrease the effects of color bias and noise. Many existing interpretable deep-learning algorithms exploit the Retinex theory as the basis of model design. However, previous Retinex-based algorithms, that consider reflected objects as ideal Lambertian ignore specular reflection in the modeling process and construct the physical constraints in image space, limiting generalization of the model. To address this issue, we preserve the specular reflection coefficient and reformulate the original physical constraints in the imaging process based on the Kubelka-Munk theory, thereby constructing constraint relationship between illumination, reflection, and detection, the so-called triple physical constraints (TPCs)theory. Based on this theory, the physical constraints are constructed in the feature space of the model to obtain the TPC network (TPCNet). Comprehensive quantitative and qualitative benchmark and ablation experiments confirm that these constraints effectively improve the performance metrics and visual quality without introducing new parameters, and demonstrate that our TPCNet outperforms other state-of-the-art methods on 10 datasets.

Paper Structure

This paper contains 12 sections, 9 equations, 4 figures, 3 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Low-Light Image Enhancement via Physical Modelling
  4. Method
  5. Triple Physical Constraint Framework
  6. Overview of the Triple Physical Constraint Network
  7. Dual-Stream Cross-Guided Transformer
  8. Experiment
  9. Datasets and Settings Details
  10. Main Results
  11. Ablation Study
  12. Conclusion

Figures (4)

  • Figure 1: Our TPCNet shows its efficient design and outstanding performance compared to recent SOTA methods, including CIDNet (HVI 6, CVPR2025), QuadPrior (Quadruple Priors 2, CVPR2024), SMG (SMG-LLIE 36, CVPR2023 ), Retinexformer 12 (ICCV2023), RQ-LLIE 37 (ICCV2023), and SNR-Aware (SNR-Net 16, CVPR2022) for Low-light Image enhancement on LOLv2-real 29 benchmarks.
  • Figure 2: (a) Overview of our proposed TPCNet, which contains a Light & Reflectivity Features Estimator (i), the Color-Association Mechanism (CAM) (ii), and the Dual-Stream Cross-Guided Transformer (iii). (b1) The Cross-Guided Attention Blocks (CGABs) are comprised of a normalization layer (LN), a Cross-Guided Multi-Head Self-Attention (CG-MSA) module, and an IEL module 6. (b2) CGAB (V) with variation in skip connection. (c1)-(c2) Detailed structure of the CG-MSA and its variation.
  • Figure 3: Comparison of the enhanced images with various SOTA methods on LOL-v2-Synthetic (top row) and VILNC-Indoor (bottom row).
  • Figure 4: Visual comparison on the MEF 40, NPE 41, LIME 11_39, DICM 38, and VV 42 datasets.