Adaptive Dual Aggregation Network with Normalizing Flows for Low-Light Image Enhancement.

Area of Science:

• Computer Vision
• Image Processing
• Artificial Intelligence

Background:

• Low-light image enhancement (LLIE) addresses visual quality issues in images captured under poor lighting.
• Current methods often rely on pixel-level errors, limiting their ability to model complex visual discrepancies.
• Deep learning and Retinex-based approaches are common but struggle with realistic error modeling.

Purpose of the Study:

• To propose a novel network, the Adaptive Dual Aggregation Network with Normalizing Flows (ADANF), for effective low-light image enhancement.
• To address the limitations of pixel-level error functions in current LLIE methods.
• To improve the modeling of real visual errors between enhanced and normally exposed images.

Main Methods:

• An adaptive dual aggregation encoder extracts illumination-robust features by analyzing global and local image properties.
• A reversible normalizing flow decoder models visual errors by mapping images to underlying data distributions.
• A gated multi-scale information transmitting module integrates encoder features into the decoder for enhanced quality.

Main Results:

• The proposed ADANF effectively enhances low-light images.
• Experiments demonstrate superior performance on both paired and unpaired datasets.
• The method successfully models real visual errors, surpassing traditional approaches.

Conclusions:

• The ADANF provides a significant advancement in low-light image enhancement.
• Normalizing flows offer a powerful tool for modeling complex visual errors in image restoration.
• The adaptive dual aggregation encoder and multi-scale information integration contribute to robust feature extraction and enhancement.