Unsupervised Learning for Low-Light Image Enhancement

Introduction

This project explores and compares two approaches for enhancing images captured in low-light conditions: a classical method based on digital image processing and a modern method using deep neural networks. The goal is to improve the visual quality and interpretability of dark images, a crucial pre-processing step for applications in computer vision, security, and automated analysis.

Both methods were evaluated using the The Dark Face dataset, and the results show significant improvements, though each approach has distinct strengths. The implemented deep learning model is Zero-DCE++, a convolutional neural network that performs enhancement without direct supervision.

Visual Results Comparison

You can insert your comparative image here, showing the results of both methods against the original images.

 

Methods Implemented

Two fundamentally different strategies were compared to address the low-light enhancement problem.

---

1. Classical Image Processing (Manual Adjustment)

This approach relies on a sequence of controlled image transformations to increase brightness, improve contrast, and reduce noise. The workflow is transparent, interpretable, and allows for fine-grained control over the final result.

The process consists of the following steps:

1. Gamma Correction: A gamma correction (${\gamma=0.5}$) is applied to amplify low-intensity values and brighten the overall image.
2. Contrast Enhancement (CLAHE): The image is converted to the LAB color space to apply Contrast Limited Adaptive Histogram Equalization (CLAHE) on the Luminance (L) channel. This improves local contrast without overexposing the image.
   • Clip Limit: 2.0
   • Tile Grid Size: (8, 8)
3. Noise Reduction: A Median Filter (${\text{kernel}=3\times3}$) is used to smooth artifacts while preserving edges.
4. RGB Conversion: Finally, the image is converted back to the RGB color space for display.

---

2. Deep Neural Network (Zero-DCE++)

This approach uses the Zero-DCE++ model, a convolutional neural network trained to enhance images without requiring paired data (i.e., dark/light versions) for supervision. The network learns to generate an enhancement map that is iteratively applied to adaptively adjust the image's luminance.

• Advantages: Its main strengths are automation and its ability to generalize across a wide variety of dark scenes.
• Training: The model was trained with specific loss functions to preserve spatial consistency, color balance, and control exposure.
• Implementation: The training was performed in PyTorch using the Adam optimizer on the Dark Face dataset. You can find the training pipeline in this repository.

Evaluation and Results

To assess the quality of the resulting images without a reference "ground truth," the following objective, no-reference evaluation metrics were used.

Evaluation Metrics

• BRISQUE (Blind/Referenceless Image Spatial Quality Evaluator): Quantifies the perceptual quality of an image.
• NIQE (Natural Image Quality Evaluator): Measures the statistical naturalness of an image.
• Entropy: Indicates the amount of information or detail present in the image.
• RMS Contrast: Measures the standard deviation of pixel intensities (global contrast).

Quantitative Comparison

The following table presents the metric results for a sample of 4 images from the dataset:

Image	Method	BRISQUE	NIQE	Entropy	RMS Contrast
1259.png	Original	44.75	5.15	2.84	15.04
	Manual Tuning	41.92	4.78	4.45	40.20
	Zero-DCE++	38.93	3.77	4.17	80.87
2454.png	Original	67.16	5.75	2.09	11.86
	Manual Tuning	48.05	5.10	3.84	32.56
	Zero-DCE++	48.40	4.57	3.59	67.82
2796.png	Original	33.89	4.42	2.87	23.79
	Manual Tuning	39.56	4.88	4.61	43.83
	Zero-DCE++	39.98	3.81	4.48	74.05
2835.png	Original	16.22	4.02	3.43	15.94
	Manual Tuning	35.48	4.37	4.91	37.83
	Zero-DCE++	11.37	2.99	5.00	52.27

Table extracted from the project's results.

Results Analysis

• Zero-DCE++ consistently achieves the highest RMS contrast and the best (lowest) NIQE scores, suggesting greater perceived naturalness and a more aggressive exposure correction.
• The manual adjustment method tends to yield higher entropy values, indicating better preservation of fine details. Visually, it produces more balanced results with a lower risk of overexposure.

Conclusion

Both methods are effective, but their suitability depends on the use case:

• Manual Adjustment is ideal for controlled environments where precise control is needed, the goal is to preserve subtle details, and computational resources are limited.
• Zero-DCE++ is superior for automated and scalable pipelines that require a robust, real-time response to diverse lighting conditions.

Code and Resources

• Language: Python
• Core Libraries: PyTorch, OpenCV, NumPy, Matplotlib, PIL
• Dataset: The Dark Face Dataset on Kaggle
• Training Pipeline: Jupyter Notebook on GitHub
• Training Pipeline: Article on GitHub

How to Cite This Project

If you use this work, please cite the repository:

@misc{Camilo2024LowLightEnhancer,
  author = {David Camilo Muñoz Garcia},
  title = {Unsupervised Learning for Low-Light Image Enhancement},
  year = {2024},
  publisher = {GitHub},
  journal = {GitHub repository},
  howpublished = {\url{https://github.com/davidcamilo0710/low_light_enhancer}}
}