CuDi: Curve Distillation for Efficient and Controllable Exposure Adjustment.

Area of Science:

• Computer Vision
• Image Processing
• Machine Learning

Background:

• Low-light image enhancement is crucial for various applications.
• Existing methods often require paired data or struggle with controllable adjustments.
• Zero-reference learning methods offer data-free training but can be slow.

Purpose of the Study:

• To develop an efficient and controllable exposure adjustment method.
• To reduce model size and inference time for low-light image enhancement.
• To enable correction of both underexposed and overexposed images with a single model.

Main Methods:

• Curve Distillation (CuDi) utilizes a novel curve distillation technique.
• Approximates iterative operations using high-order curve tangent lines for speed.
• Employs a self-supervised spatial exposure control loss for region-specific adjustments.
• Leverages an input condition exposure map for global or local control.

Main Results:

• Achieved significant speed-up in inference and reduction in model size.
• Demonstrated effective correction for both underexposed and overexposed images.
• Enabled flexible, controllable exposure adjustment guided by exposure maps.

Conclusions:

• CuDi provides a fast, robust, and flexible solution for image exposure adjustment.
• Outperforms state-of-the-art methods in real-world scenarios.
• Offers a valuable tool for image enhancement without requiring paired training data.