DRLIE: Flexible Low-Light Image Enhancement via Disentangled Representations

Area of Science:

• Computer Vision
• Image Processing

Background:

• Low-light image enhancement (LIME) is crucial for improving image quality in underexposed conditions.
• Existing LIME methods often lack controllability, leading to unpredictable results.

Purpose of the Study:

• To develop a flexible and controllable framework for low-light image enhancement.
• To enable users to guide the enhancement process using reference images.

Main Methods:

• A novel framework models images by decoupling content and exposure attributes.
• Content and attribute encoders disentangle these components.
• A generator reconstructs the enhanced image using low-light content and guide image exposure.

Main Results:

• The proposed method significantly outperforms state-of-the-art LIME techniques on public datasets.
• User-specified guide images allow for personalized image enhancement.
• Demonstrated superior practicability and controllability in low-light conditions.

Conclusions:

• The developed framework offers a controllable and effective solution for low-light image enhancement.
• Information decoupling provides a robust approach for manipulating image attributes.
• The method enhances practical usability by incorporating user preferences through guide images.