Towards a Flexible Semantic Guided Model for Single Image Enhancement and Restoration

Area of Science:

• Computer Vision
• Image Processing
• Artificial Intelligence

Background:

• Low-light image enhancement (LLIE) aims to improve image quality in dark conditions.
• Existing LLIE methods often apply uniform enhancement, neglecting regional semantic details and leading to color distortion.
• Incorporating semantic information is crucial for accurate and natural low-light image enhancement.

Purpose of the Study:

• To propose a novel semantic-aware knowledge-guided framework (SKF) for low-light image enhancement.
• To develop a generalizable framework that integrates semantic priors into LLIE models.
• To introduce SKF++ with additional techniques for enhanced intra-class recovery and contrast adjustment.

Main Methods:

• Developed a semantic-aware embedding module to integrate semantic priors into feature representations.
• Introduced a semantic-guided color histogram loss for preserving color consistency.
• Utilized a semantic-guided adversarial loss for generating natural textures.
• Proposed SKF++ with an extra convolutional branch and equalization-based histogram transformation.

Main Results:

• Models equipped with SKF/SKF++ significantly outperformed baseline methods on LLIE benchmarks.
• The proposed framework demonstrated strong generalization across different models and scenes.
• SKF/SKF++ showed potential benefits for downstream tasks like face detection and semantic segmentation in low-light environments.

Conclusions:

• The semantic-aware knowledge-guided framework (SKF) effectively enhances low-light images by leveraging semantic information.
• SKF++ offers further improvements in image quality and dynamic range.
• The proposed methods provide a robust and generalizable solution for low-light image enhancement and related applications.