A Comprehensive Review of Low-Light Image Enhancement Based on Deep Learning

0
0
0

A Comprehensive Review of Low-Light Image Enhancement Based on Deep Learning

Introduction

Images serve as crucial mediums for information transmission in various fields, including surveillance, autonomous driving, and medical imaging. However, images captured in low-light conditions often suffer from poor brightness, high noise levels, and low contrast, significantly degrading their quality and usability. These issues can negatively impact downstream computer vision tasks such as object detection, facial recognition, and scene understanding. Enhancing low-light images is therefore essential to improve their interpretability and applicability in real-world scenarios.

Traditionally, low-light image enhancement has been approached through hardware and software solutions. Hardware-based methods involve improving camera sensor sensitivity and optimizing sensor-processor collaboration, but these approaches are often costly. Consequently, software-based solutions have become the dominant strategy, with traditional methods relying on pixel-value mapping techniques such as histogram equalization and gamma correction. However, these methods often fail to adequately enhance image quality, leading to noise amplification and loss of fine details.

In recent years, deep learning has revolutionized low-light image enhancement by leveraging powerful neural networks to learn complex mappings between low-light and well-lit images. Deep learning-based approaches can be broadly categorized into supervised and unsupervised methods. Supervised methods rely on paired datasets of low-light and normal-light images to train models, while unsupervised methods explore intrinsic relationships between low-light and normal-light images without requiring paired training data. This article provides a comprehensive review of traditional and deep learning-based low-light image enhancement techniques, discussing their advantages, limitations, and future research directions.

Traditional Low-Light Image Enhancement Methods

Traditional methods for low-light image enhancement can be classified into two main categories: distribution mapping-based and model optimization-based approaches.

Distribution Mapping-Based Methods

Distribution mapping techniques adjust pixel intensity distributions to improve image visibility. Common methods include:

  1. Histogram Equalization – This technique redistributes pixel intensities to enhance global or local contrast. While simple and effective, it often amplifies noise in dark regions and loses fine details.
  2. Gamma Correction – A nonlinear transformation that adjusts image brightness and contrast. The effectiveness of gamma correction depends on selecting an appropriate gamma value, which requires prior knowledge. Improper gamma values can lead to over-correction.
  3. Tone Mapping – This method improves dynamic range by adjusting brightness and contrast. Global tone mapping applies uniform adjustments, while local tone mapping adapts to different regions. However, tone mapping can introduce halo artifacts and color distortions when neighboring pixel intensities vary significantly.

Model Optimization-Based Methods

These methods rely on mathematical models to describe image characteristics and degradation processes. Key approaches include:

  1. Retinex Algorithm – Based on the separation of illumination and reflectance components, Retinex theory estimates the illumination component to enhance low-light images. While effective in brightness improvement, it may introduce halo artifacts.
  2. Dark Channel Prior – This approach assumes that natural images contain dark regions with minimal illumination. By estimating global atmospheric light, it recovers image details. However, it may over-enhance images and amplify noise.

Limitations of Traditional Methods

Despite their contributions, traditional methods have several drawbacks:
• Noise amplification in dark regions.

• Loss of texture and fine details.

• Limited dynamic range expansion, leading to highlight clipping.

Deep Learning-Based Low-Light Image Enhancement

Deep learning has become the dominant approach in low-light image enhancement due to its ability to learn complex mappings between low-light and normal-light images. These methods can be divided into supervised and unsupervised learning paradigms.

Supervised Learning Methods

Supervised methods require paired datasets of low-light and normal-light images for training. Some notable approaches include:

  1. LLNet – One of the earliest deep learning-based methods, LLNet uses a stacked sparse denoising autoencoder to enhance brightness and reduce noise. However, it suffers from artifacts and color distortions.
  2. Retinex-Net – Inspired by Retinex theory, this model decomposes images into illumination and reflectance components. While effective in brightness enhancement, it may produce artifacts and color inconsistencies.
  3. KinD and KinD++ – These models improve upon Retinex-Net by incorporating multi-scale illumination attention and data alignment strategies, enhancing visual quality but sometimes causing overexposure.
  4. SICE Network – This model uses a dual-branch architecture (U-Net and residual network) to extract features and preserve details. It employs multiple loss functions to reduce artifacts and noise.
  5. Transformer-Based Models – Recent advancements integrate Retinex theory with Transformer architectures, achieving superior performance in brightness and detail recovery.

Unsupervised Learning Methods

Unsupervised methods eliminate the need for paired datasets, making them more flexible for real-world applications. Key approaches include:

  1. Zero-DCE and Zero-DCE++ – These models use zero-reference deep curve estimation to iteratively adjust local brightness. Zero-DCE++ reduces computational complexity, making it suitable for mobile devices.
  2. EnlightenGAN – A generative adversarial network (GAN) that employs global-local discriminators to generate realistic enhanced images. However, it may introduce noise and color shifts.
  3. LE-GAN – This GAN-based model incorporates spatial and global illumination attention mechanisms to address overexposure issues.
  4. MAGAN – Uses mixed attention modules to enhance low-light images while suppressing noise.

Loss Functions in Deep Learning

Loss functions play a critical role in training deep learning models for low-light enhancement. Commonly used loss functions include:
• Mean Absolute Loss (L1 Loss) – Measures pixel-wise differences between enhanced and reference images.

• Structural Similarity Index Loss (SSIM Loss) – Ensures structural consistency between enhanced and reference images.

• Perceptual Loss – Uses pre-trained networks (e.g., VGG) to maintain high-level feature similarity.

• Adversarial Loss – Improves realism by training a discriminator to distinguish between real and enhanced images.

Datasets and Evaluation Metrics

Datasets

Several datasets are commonly used for training and evaluating low-light enhancement models:
• LOL – Contains 500 paired low-light and normal-light images.

• SID – Includes short-exposure and corresponding long-exposure images for training.

• MIT-Adobe FiveK – Provides professionally retouched images for reference.

• ExDARK – A large-scale dataset of low-light images across various scenes.

Evaluation Metrics

Objective evaluation metrics include:
• Peak Signal-to-Noise Ratio (PSNR) – Measures pixel-level similarity between enhanced and reference images. Higher values indicate better quality.

• Structural Similarity Index (SSIM) – Evaluates structural similarity, with values closer to 1 indicating better preservation of image structure.

• Information Entropy (IE) – Quantifies the amount of information in an image, with higher values indicating richer details.

Performance Comparison

Traditional Methods

Retinex-based methods achieve the highest IE values (6.871), outperforming histogram equalization (5.406) and gamma correction (6.521). However, traditional methods generally struggle with noise and detail preservation.

Deep Learning Methods

Transformer-based models, such as RetinexFormer, achieve the highest PSNR (28.4869) and SSIM (0.9374) scores, demonstrating superior performance in brightness and detail recovery. Unsupervised methods like Zero-DCE++ offer practical solutions with lower computational costs but may introduce artifacts.

Challenges and Future Directions

Despite significant progress, several challenges remain:

  1. Robustness and Adaptability – Supervised methods rely on paired datasets, limiting their generalization. Zero-shot learning and unsupervised methods show promise in addressing this issue.
  2. Lightweight Model Design – Many deep learning models are computationally intensive. Lightweight architectures (e.g., MobileNet-based models) are needed for real-time applications.
  3. Task-Specific Evaluation Metrics – Current metrics (PSNR, SSIM) may not fully capture perceptual quality. Developing specialized metrics for noise, artifacts, and color fidelity is essential.

Conclusion

Deep learning has significantly advanced low-light image enhancement, offering superior performance over traditional methods. Supervised models excel in brightness and detail recovery, while unsupervised methods provide flexibility in real-world scenarios. Future research should focus on improving model robustness, reducing computational complexity, and developing better evaluation metrics to further enhance low-light image quality.

doi.org/10.19734/j.issn.1001-3695.2024.06.0176

Was this helpful?

0 / 0