NTIRE 2025 Challenge on Low Light Image Enhancement: Methods and Results

The evolution of low-light image enhancement has transitioned from classical algorithms like histogram equalization and Retinex to deep learning models such as CNNs and Transformers. Early methods improved brightness but often amplified noise and distorted colors. CNN-based models like U-Net, ResNet, and multi-scale architectures significantly advanced the field, enabling better detail preservation and noise suppression. Recent breakthroughs include Transformer-based models like Restormer and structure-guided networks like ESDNet, which excel in modeling long-range dependencies and multi-scale features. Despite these advances, challenges persist in handling extreme darkness, non-uniform illumination, and computational efficiency, especially for high-resolution images. The NTIRE 2025 challenge builds on this foundation, encouraging innovative fusion of these techniques to push performance further.

The core challenge in low-light enhancement is achieving simultaneous brightness increase, noise reduction, and detail preservation under complex lighting conditions. Existing models often struggle with noise amplification, color distortion, and loss of fine details, especially in extreme darkness or non-uniform illumination. Moreover, balancing computational efficiency with high-quality output remains difficult, limiting real-time deployment. Addressing these issues requires novel architectures capable of multi-scale feature fusion, robust noise suppression, and adaptive scene understanding. The problem is compounded by the diversity of real-world low-light scenarios, demanding models that are both highly accurate and computationally feasible.

Key innovations include: 1) Multi-model fusion strategies like FusionNet, combining local and global features for robustness; 2) Structure-guided Transformer modules (SG-LLIE) that incorporate structural priors and long-range dependencies; 3) Frequency domain feature integration to enhance noise suppression and color fidelity; 4) Multi-loss optimization combining Charbonnier, perceptual, and multi-scale SSIM losses for balanced training. These innovations collectively enable models to outperform previous state-of-the-art methods, especially in challenging scenarios, by leveraging complementary strengths of different architectures and training paradigms.

• �� Data: Utilized NTIRE 2025 dataset with diverse low-light scenes, including indoor, outdoor, night, and backlit images. • Architecture: Employed multi-model fusion (FusionNet, SG-LLIE, Restormer), each with encoder-decoder structures, multi-scale feature extraction, and Transformer modules. • Feature Fusion: Integrated long-range dependencies via structure-guided Transformers, combined frequency domain features, and employed multi-scale residual blocks. • Losses: Used Charbonnier for pixel accuracy, perceptual loss (VGG-based) for perceptual quality, multi-scale SSIM for structural fidelity, and color loss for color accuracy. • Training: Multi-stage, with data augmentation, self-ensembling, and transfer learning, trained on large datasets with Adam optimizer, cyclic cosine annealing, and early stopping. • Inference: Employed model ensembling, multi-scale testing, and image cropping for high-resolution images to balance accuracy and efficiency.

Experiments involved 219 training, 46 validation, and 30 test scenes, evaluated via PSNR, SSIM, LPIPS, NIQE. Ablation studies confirmed the importance of multi-model fusion, structure-guided modules, and frequency features. Hyperparameters like learning rate, batch size, and training epochs were optimized through grid search. Results showed that fusion models outperform single architectures by 2-3dB in PSNR, with significant improvements in visual quality, especially in night and backlit scenes. Cross-scenario tests validated robustness, and ablation highlighted the contribution of each component.

The best models achieved PSNR of 26.35, SSIM of 0.858, LPIPS of 0.128, NIQE of 10.95, outperforming baseline methods by substantial margins. Fusion strategies and multi-scale features were critical, with structure-guided Transformers providing global context. Ablation confirmed that frequency domain features and multi-loss training significantly enhanced noise suppression and detail preservation. The models demonstrated strong generalization across diverse low-light scenarios, setting new benchmarks for the field.

These models are applicable in night photography, surveillance, autonomous driving, and augmented reality, where real-time, high-quality low-light enhancement is crucial. They require high-performance hardware for deployment but promise substantial improvements in visibility and safety. Future integration with scene understanding and adaptive processing will broaden their industrial impact, enabling smarter, more resilient vision systems.

High computational cost and slow inference limit real-time deployment on edge devices. Performance in extremely noisy or ultra-high-resolution images needs further improvement. The models' reliance on large annotated datasets poses challenges for resource-constrained environments. Future work should focus on lightweight architectures, adaptive scene-aware modules, and multi-modal data fusion to address these limitations.