Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Unified Low-Light Traffic Image Enhancement via Multi-Stage Illumination Recovery and Adaptive Noise Suppression

Siddiqua Namrah

TL;DR

The paper tackles the challenge of reliable perception in nighttime traffic imagery by proposing a fully unsupervised, multi-stage LLIE framework that decomposes images into illumination and reflectance via Retinex-inspired networks and refines them with Channel-Guidance, Color Enhancement, and Over-Exposure Correction modules. It introduces a comprehensive loss design—Projection, Consistency, Retinex, and Perceptual losses—to learn physically valid decomposition and perceptually faithful enhancements without ground-truth pairs. Across BDD100K-Night, LoLI-Street, and MVLT, the method achieves state-of-the-art or near-state-of-the-art performance in full-reference and no-reference metrics, while delivering visually coherent results suitable for real-time traffic applications. The approach significantly improves visibility, texture preservation, and downstream perception reliability in real-world low-light urban environments, with potential for video extension and edge-device deployment.

Abstract

Enhancing low-light traffic images is crucial for reliable perception in autonomous driving, intelligent transportation, and urban surveillance systems. Nighttime and dimly lit traffic scenes often suffer from poor visibility due to low illumination, noise, motion blur, non-uniform lighting, and glare from vehicle headlights or street lamps, which hinder tasks such as object detection and scene understanding. To address these challenges, we propose a fully unsupervised multi-stage deep learning framework for low-light traffic image enhancement. The model decomposes images into illumination and reflectance components, progressively refined by three specialized modules: (1) Illumination Adaptation, for global and local brightness correction; (2) Reflectance Restoration, for noise suppression and structural detail recovery using spatial-channel attention; and (3) Over-Exposure Compensation, for reconstructing saturated regions and balancing scene luminance. The network is trained using self-supervised reconstruction, reflectance smoothness, perceptual consistency, and domain-aware regularization losses, eliminating the need for paired ground-truth images. Experiments on general and traffic-specific datasets demonstrate superior performance over state-of-the-art methods in both quantitative metrics (PSNR, SSIM, LPIPS, NIQE) and qualitative visual quality. Our approach enhances visibility, preserves structure, and improves downstream perception reliability in real-world low-light traffic scenarios.

Unified Low-Light Traffic Image Enhancement via Multi-Stage Illumination Recovery and Adaptive Noise Suppression

TL;DR

The paper tackles the challenge of reliable perception in nighttime traffic imagery by proposing a fully unsupervised, multi-stage LLIE framework that decomposes images into illumination and reflectance via Retinex-inspired networks and refines them with Channel-Guidance, Color Enhancement, and Over-Exposure Correction modules. It introduces a comprehensive loss design—Projection, Consistency, Retinex, and Perceptual losses—to learn physically valid decomposition and perceptually faithful enhancements without ground-truth pairs. Across BDD100K-Night, LoLI-Street, and MVLT, the method achieves state-of-the-art or near-state-of-the-art performance in full-reference and no-reference metrics, while delivering visually coherent results suitable for real-time traffic applications. The approach significantly improves visibility, texture preservation, and downstream perception reliability in real-world low-light urban environments, with potential for video extension and edge-device deployment.

Abstract

Enhancing low-light traffic images is crucial for reliable perception in autonomous driving, intelligent transportation, and urban surveillance systems. Nighttime and dimly lit traffic scenes often suffer from poor visibility due to low illumination, noise, motion blur, non-uniform lighting, and glare from vehicle headlights or street lamps, which hinder tasks such as object detection and scene understanding. To address these challenges, we propose a fully unsupervised multi-stage deep learning framework for low-light traffic image enhancement. The model decomposes images into illumination and reflectance components, progressively refined by three specialized modules: (1) Illumination Adaptation, for global and local brightness correction; (2) Reflectance Restoration, for noise suppression and structural detail recovery using spatial-channel attention; and (3) Over-Exposure Compensation, for reconstructing saturated regions and balancing scene luminance. The network is trained using self-supervised reconstruction, reflectance smoothness, perceptual consistency, and domain-aware regularization losses, eliminating the need for paired ground-truth images. Experiments on general and traffic-specific datasets demonstrate superior performance over state-of-the-art methods in both quantitative metrics (PSNR, SSIM, LPIPS, NIQE) and qualitative visual quality. Our approach enhances visibility, preserves structure, and improves downstream perception reliability in real-world low-light traffic scenarios.

Paper Structure

This paper contains 30 sections, 7 equations, 5 figures, 4 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Classical Image Enhancement Methods
  4. Deep Learning for LLIE
  5. Traffic-Centric LLIE Approaches
  6. Proposed Method
  7. Retinex-Based Image Decomposition
  8. Channel-Guidance Module
  9. Color Enhancement Module
  10. Over-Exposure Correction Module
  11. Loss Function Design
  12. Projection Loss for Stable Normalization
  13. Consistency Loss for Reflectance Invariance
  14. Retinex Loss for Physically Valid Decomposition
  15. Perceptual Loss for Semantic and Color Fidelity
  16. ...and 15 more sections

Figures (5)

  • Figure 1: Using the BDD100K Dataset, different low-light image enhancement (LIE) methods are visually compared. The initial low-light input picture is shown in (a). (b–i) show the results produced by a number of cutting-edge enhancing techniques. The outcome of our suggested approach, which yields considerably superior enhancement, is shown in (j). These comparisons (b–i) highlight the advantages and disadvantages of each strategy; our solution (j) successfully recovers details and improves color fidelity in low light.
  • Figure 2: Using the LoLI-Street dataset, different low-light image enhancement (LIE) techniques are visually compared. The first low-light input is seen in (a). (b–j) show improvement outcomes from cutting-edge techniques. (k) displays the result of our suggested approach, which exhibits excellent color accuracy and detail restoration. The equivalent ground truth is (l). These contrasts show the advantages and disadvantages of each strategy in difficult low-light situations..
  • Figure 3: Using the MVLT Dataset, different low-light image enhancement (LIE) methods are visually compared. The original low-light input picture is shown in (a). (b–j) show the results produced by several cutting-edge enhancing techniques. The outcome of our suggested approach, which produces considerably superior enhancement, is shown in (k). (l) Shows the picture of the ground truth.
  • Figure 4: Ablation studies on the BDD100k, MVLT, and LoLI-Street Datasets are visually compared. (a) Without the OEC Module; (b) Without the CG Module; (c) Without the CE Module; and (d) The output of our whole model. The visual outcomes show that the whole model offers the most accurate improvement, with each part being essential to raising the output quality..
  • Figure :