Learning to Enhance Low-Light Image via Zero-Reference Deep Curve Estimation.

Area of Science:

• Computer Vision
• Deep Learning
• Image Processing

Background:

• Low-light image enhancement is crucial for various applications.
• Existing methods often require paired or unpaired training data, limiting their applicability.
• Deep learning approaches have shown promise but still face challenges in generalization and data requirements.

Purpose of the Study:

• To introduce a novel zero-reference deep curve estimation method for low-light image enhancement.
• To develop a lightweight deep network capable of estimating pixel-wise curves for dynamic range adjustment.
• To achieve effective image enhancement without relying on any reference images during training.

Main Methods:

• Formulating light enhancement as image-specific curve estimation using a deep network (DCE-Net).
• Designing curve estimation considering pixel value range, monotonicity, and differentiability.
• Utilizing non-reference loss functions to implicitly measure and drive enhancement quality.

Main Results:

• Zero-DCE demonstrates effective generalization across diverse lighting conditions.
• An accelerated version, Zero-DCE++, achieves fast inference speeds with minimal parameters.
• Qualitative and quantitative experiments show superior performance compared to state-of-the-art methods.

Conclusions:

• Zero-DCE offers a powerful and data-efficient solution for low-light image enhancement.
• The method's simplicity and efficiency make it suitable for real-world applications.
• Potential benefits for tasks like face detection in dark environments are highlighted.