Visual Perception Model for Rapid and Adaptive Low-light Image Enhancement

Xiaoxiao Li, Xiaopeng Guo, Liye Mei, Mingyu Shang, Jie Gao, Maojing Shu, Xiang WangarXiv 2020 · arXiv

Abstract

Low-light image enhancement is critical for computer vision tasks where environmental lighting is insufficient. Traditional methods based on Retinex theory often struggle with over-saturation or persistent darkness due to a reliance on single light-intensity estimates. This paper proposes a novel Visual Perception (VP) model that decomposes light into intensity and spatial distribution. By employing a cycle operation with an adaptive comparator, the framework achieves high-quality image restoration with superior computational speed compared to state-of-the-art methods.

Introduction

The Human Vision System (HVS) is naturally limited in low-light environments, leading to images with poor contrast and obscured details. For decades, the Retinex theory—which models an image as the product of illumination and reflectance—has served as the foundation for enhancement algorithms. However, this model often oversimplifies the lighting environment, leading to "visual photosensitive imbalances" where shadows remain too dark or highlights become "blown out."

Furthermore, many existing enhancement techniques lack adaptivity, requiring manual parameter tuning that fails across diverse lighting conditions. To address these issues, this research introduces a Visual Perception (VP) model designed to bridge the gap between mathematical modeling and actual human visual perception, creating a rapid, automated enhancement pipeline.

The Visual Perception Model

The proposed framework moves beyond simple light-intensity estimation by modeling how the human eye perceives light distribution. The process is divided into four primary stages:

  1. Preprocessing: The input image is converted to the HSI (Hue, Saturation, Intensity) color space. The intensity channel is smoothed to mitigate noise that could interfere with perception modeling.
  2. VP Information Modeling: Unlike the standard Retinex model, the VP model quantifies four distinct features: luminance, obscuration (darkness), and the spatial distribution of both. This allows the algorithm to understand the context of light within the scene.
  3. Adaptive Estimation: To maintain mathematical efficiency, the authors developed adaptive $\beta$ and $\gamma$ functions. The $\beta$ function quantifies dark-area imbalance, while $\gamma$ assists in reflectance estimation, automatically adjusting weights based on the specific input image.
  4. Cycle Operation and Reconstruction: The system generates multiple enhanced versions through a feedback loop. An adaptive comparator evaluates these results against a threshold to select the version that best balances brightness, detail, and naturalness.
Comparison between the traditional Retinex model (a.1) and the proposed Visual Perception (VP) model (a.2), highlighting the inclusion of spatial distribution and perception refinement.

Experimental Results

The researchers evaluated the VP model using three benchmarks: the Retinex Dataset (NASA-retouched), the Kodak Dataset (lossless color), and the DICM Dataset (real-world low-light).

Visual and Quantitative Performance

The VP model was compared against six state-of-the-art (SOTA) methods, including KIND, LIME, and NPE. Visually, the VP model avoided common artifacts such as the "oil painting" look or excessive blurring (atomization). Quantitatively, it achieved leading scores in Peak Signal-to-Noise Ratio (PSNR), Structural Similarity (SSIM), and Perception-based Image Quality Evaluator (PIQE).

Visual comparison of enhancement results. The VP model (d.1) demonstrates superior balance in brightness and detail preservation compared to the original low-light input (a.1) and intermediate results.

Computational Efficiency

A key advantage of the VP model is its speed. By condensing complex parameters into integrated functions, the model processes a $500 \times 1000$ image in approximately 0.3 to 0.5 seconds. In contrast, competing methods like NPE can take upwards of 12 seconds for the same task, making the VP model highly suitable for practical, real-time applications.

Conclusion

The Visual Perception model offers a significant advancement over Retinex-based methods by incorporating spatial light distribution and an adaptive feedback loop. While the model excels at static image enhancement, future work will focus on extending this framework to video streams, where temporal consistency is a primary challenge, and further improving noise handling in extremely dark environments.

Summary updated

BibTeX
@misc{li2020visual,
  author = {Xiaoxiao Li and Xiaopeng Guo and Liye Mei and Mingyu Shang and Jie Gao and Maojing Shu and Xiang Wang},
  title = {{Visual Perception Model for Rapid and Adaptive Low-light Image Enhancement}},
  year = {2020},
  eprint = {2005.07343},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2005.07343},
}
Figure