Rethinking progressive low-light image enhancement: A frequency-aware tripartite multi-scale network.

Area of Science:

• Computer Vision
• Image Processing
• Artificial Intelligence

Background:

• Low-light images suffer from contrast reduction, color distortion, and noise.
• Progressive networks show promise for low-light image enhancement but often lack effective multi-scale feature fusion.

Purpose of the Study:

• To propose a novel Progressive Tripartite Multi-Scale Network (PTMSNet) for effective low-light image enhancement.
• To address the limitations of single-scale fusion strategies in existing progressive methods.

Main Methods:

• Designed a Parallel Hybrid Module (PHM) integrating Transformer and Convolutional Neural Network (CNN) branches for complementary feature learning.
• Introduced a Frequency-Aware Fusion Module (FAFM) for accurate Signal-to-Noise Ratio (SNR) mapping using frequency-domain information.
• Employed a joint loss function incorporating pixel, structural, and color constraints.

Main Results:

• PTMSNet significantly outperforms state-of-the-art (SOTA) methods on benchmark datasets.
• Demonstrated superior performance on the ExDark dataset, indicating robust low-light detection.
• Showcased practical applicability in real-world scenarios with unstable illumination.

Conclusions:

• The proposed PTMSNet effectively enhances low-light images by leveraging multi-scale feature interactions.
• The network achieves superior performance due to its innovative hybrid and frequency-aware fusion modules.
• PTMSNet offers robust generalization capabilities for real-world applications requiring high-quality low-light imaging.