Open AccessMCATD: Multi-scale Contextual Attention Transformer Diffusion for Unsupervised Low-light Image Enhancement
Author(s) -
Cheng Da,
Yongsheng Qian,
Junwei Zeng,
Xuting Wei,
Futao Zhang
Publication year - 2025
Publication title -
ieee access
Language(s) - English
Resource type - Magazines
SCImago Journal Rank - 0.587
H-Index - 127
eISSN - 2169-3536
DOI - 10.1109/access.2025.3573171
Subject(s) - aerospace , bioengineering , communication, networking and broadcast technologies , components, circuits, devices and systems , computing and processing , engineered materials, dielectrics and plasmas , engineering profession , fields, waves and electromagnetics , general topics for engineers , geoscience , nuclear engineering , photonics and electrooptics , power, energy and industry applications , robotics and control systems , signal processing and analysis , transportation
Low-light image enhancement (LLIE) remains a challenging task due to the complex degradation patterns in images captured under insufficient illumination, including non-linear intensity mappings, spatially-varying noise distributions, and content-dependent color distortions. Despite significant advances, existing methods struggle with three fundamental challenges: (1) difficulty in simultaneously preserving structural details while reducing noise, (2) limited generalization across diverse lighting conditions and scene types, and (3) computational inefficiency when processing complex natural scenes. While recent diffusion-based methods have shown promise, they often struggle with generalization and require paired training data. We propose MCATD, a novel unsupervised framework that integrates adaptive sampling, multi-scale feature extraction, and dynamic enhancement capabilities into diffusion models for LLIE. The framework consists of three key components: (1) a Dynamic Adaptive Diffusion Sampling (DADS) strategy that adjusts sampling steps based on image complexity, (2) a Multi-scale Contextual Attention Transformer (MCAT) network that captures features at different scales with attention mechanisms, and (3) a Multi-scale Dynamic Structure-Preserving (MDSP) loss that preserves image structure while optimizing perceptual quality. Experimental results on multiple benchmarks demonstrate that our method outperforms state-of-the-art unsupervised approaches and achieves comparable performance to supervised methods while maintaining better generalization ability. Furthermore, ablation studies validate the effectiveness of each proposed component. The proposed framework not only advances the field of unsupervised LLIE but also provides insights into leveraging diffusion models for image restoration tasks.