Dimma: Semi-Supervised Low-Light Image Enhancement with Adaptive Dimming.

Area of Science:

• Computer Vision
• Image Processing
• Machine Learning

Background:

• Enhancing low-light images naturally is difficult due to camera variations and lack of ground-truth data.
• Existing methods often require extensive labeled datasets for training.

Purpose of the Study:

• To develop a semi-supervised method (Dimma) for natural color enhancement of low-light images.
• To create a camera-agnostic approach requiring minimal training data.

Main Methods:

• Utilizing a convolutional mixture density network to model camera-specific noise in dark images.
• Employing a conditional UNet architecture incorporating user-defined lightness values.
• Training on a small set of real image pairs captured under extreme lighting.

Main Results:

• Dimma effectively enhances low-light images, preserving natural colors.
• The method demonstrates competitive performance against fully supervised state-of-the-art techniques.
• Achieves robust results with limited training data.

Conclusions:

• Dimma offers an efficient and effective solution for low-light image enhancement.
• The semi-supervised approach reduces the need for large, specialized datasets.
• Provides a practical method for improving image quality across various cameras.