Deep Guided Exposure Correction with Knowledge Distillation

Area of Science:

• Computer Vision
• Image Processing
• Deep Learning

Background:

• Image exposure problems like over- and under-exposure significantly degrade visual quality.
• Existing methods often focus on sRGB images and neglect valuable channel information, limiting their effectiveness.

Purpose of the Study:

• To develop a unified deep learning network for simultaneous RAW image exposure correction.
• To leverage channel-specific information for improved exposure correction accuracy.

Main Methods:

• Propose a Deep Guided network with Knowledge Distillation (DGKD) for RAW image exposure correction.
• Utilize blue/red channels for over-exposure guidance and green channel for under-exposure guidance, based on color sensitivity.
• Employ knowledge distillation to integrate individual over- and under-exposure correction networks into a single unified network.

Main Results:

• The DGKD method demonstrates superior performance compared to state-of-the-art techniques in both quantitative metrics and visual quality.
• Achieved a peak signal-to-noise ratio of 24.653 dB and a structural similarity index of 0.8182 on a dedicated RAW image dataset.

Conclusions:

• The proposed DGKD network offers an effective solution for simultaneous RAW image over- and under-exposure correction.
• Leveraging channel information and knowledge distillation enhances the robustness and performance of exposure correction models.