Towards Scale-Aware Low-Light Enhancement via Structure-Guided Transformer Design

2025-04-18

AI Summary

To address the severe ill-posedness and semantic feature degradation in extremely low-light image enhancement (LLIE), this paper proposes SG-LLIE, a structure-guided CNN-Transformer hybrid framework. Methodologically, we introduce the first illumination-invariant edge-driven Structure-Guided Transformer Block (SGTB), eliminating reliance on pre-trained illumination or semantic maps; and design a Hierarchical Structure-Guided Feature Extractor (HSGFE) that explicitly embeds edge-structural priors into feature modulation. The architecture integrates a U-Net encoder-decoder backbone, multi-scale feature fusion, and structure-aware self-attention. Evaluated on multiple mainstream LLIE benchmarks, SG-LLIE achieves state-of-the-art performance, ranking second in the NTIRE 2025 Low-Light Image Enhancement Challenge. Quantitative metrics (e.g., PSNR, SSIM) and visual quality both show significant improvements over prior methods.

Technology Category

• Machine Learning: Large Multimodal Models (LMMs)
• Computer Vision: Low Level & Physics-based Vision
• Search and Optimization: Learning to Search

Application Category

• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs
• Search and Retrieval-Augmented AI: Large language models for search
• Graph Algorithms and Modeling for the Web: Foundation models and LLMs for Web-related graphs

Abstract

Current Low-light Image Enhancement (LLIE) techniques predominantly rely on either direct Low-Light (LL) to Normal-Light (NL) mappings or guidance from semantic features or illumination maps. Nonetheless, the intrinsic ill-posedness of LLIE and the difficulty in retrieving robust semantics from heavily corrupted images hinder their effectiveness in extremely low-light environments. To tackle this challenge, we present SG-LLIE, a new multi-scale CNN-Transformer hybrid framework guided by structure priors. Different from employing pre-trained models for the extraction of semantics or illumination maps, we choose to extract robust structure priors based on illumination-invariant edge detectors. Moreover, we develop a CNN-Transformer Hybrid Structure-Guided Feature Extractor (HSGFE) module at each scale with in the UNet encoder-decoder architecture. Besides the CNN blocks which excels in multi-scale feature extraction and fusion, we introduce a Structure-Guided Transformer Block (SGTB) in each HSGFE that incorporates structural priors to modulate the enhancement process. Extensive experiments show that our method achieves state-of-the-art performance on several LLIE benchmarks in both quantitative metrics and visual quality. Our solution ranks second in the NTIRE 2025 Low-Light Enhancement Challenge. Code is released at https://github.com/minyan8/imagine.

Problem

• Enhancing extremely low-light images robustly
• Overcoming ill-posedness in low-light enhancement
• Integrating structure priors without pre-trained models

Innovation

• CNN-Transformer hybrid framework with structure priors
• Illumination-invariant edge detectors for robust priors
• Structure-Guided Transformer Block modulates enhancement