EnlightenGAN: Deep Light Enhancement Without Paired Supervision

Area of Science:

• Computer Vision
• Artificial Intelligence
• Image Processing

Background:

• Deep learning excels at image restoration but often requires paired training data.
• Acquiring paired low-light and normal-light images of the same scene is practically challenging.
• Existing methods struggle with the lack of paired data for low-light enhancement.

Purpose of the Study:

• To develop an effective unsupervised method for low-light image enhancement.
• To address the challenge of limited paired training data in image restoration tasks.
• To propose a generative adversarial network (GAN) that performs well without ground truth pairs.

Main Methods:

• Proposed EnlightenGAN, an unsupervised generative adversarial network architecture.
• Introduced innovations including a global-local discriminator, self-regularized perceptual loss fusion, and attention mechanisms.
• Employed self-regularization using input image information instead of paired ground truth data.

Main Results:

• EnlightenGAN demonstrates highly competitive performance in low-light image enhancement.
• The unsupervised approach generalizes effectively to various real-world low-light images.
• Outperformed recent methods in visual quality and subjective user studies.

Conclusions:

• Unsupervised learning is a viable and effective strategy for low-light image enhancement.
• EnlightenGAN offers a flexible solution adaptable to diverse image domains.
• The proposed method provides a strong baseline for future research in unpaired image restoration.