Boosting Diffusion Networks with Deep External Context-Aware Encoders for Low-Light Image Enhancement.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Image Processing

Background:

• Low-light image enhancement (LLIE) is challenging due to complex, widespread degradations.
• Existing diffusion models struggle with global context modeling, leading to high computational costs.

Purpose of the Study:

• To develop an efficient diffusion-based LLIE method that enhances long-range context modeling.
• To reduce the computational overhead associated with global module integration in diffusion backbones.

Main Methods:

• Proposed ECA-Diff, a diffusion framework with an External Context-Aware Encoder (ECAE).
• Utilized a latent-space context network with hybrid Transformer-Convolution blocks for feature extraction.
• Implemented a CIELAB-space Luminance-Adaptive Chromaticity Loss for training regularization.

Main Results:

• ECA-Diff outperformed state-of-the-art LLIE methods on paired and unpaired benchmarks.
• Achieved superior performance in both full-reference (PSNR, SSIM, LPIPS) and no-reference (NIQE, BRISQUE) metrics.
• The external context path added only modest computational overhead.

Conclusions:

• Decoupling global context estimation from iterative denoising effectively boosts diffusion-based LLIE.
• ECA-Diff offers a generalizable compute-once conditioning paradigm for low-level image restoration tasks.