- The paper introduces a two-stage pipeline combining transmittance correction and STAR-YUV structure-texture decomposition for comprehensive nighttime dehazing.
- It demonstrates improved color fidelity with a 20.1% better CIEDE2000 score and enhanced detail preservation validated via PSNR (17.024 dB) and SSIM (0.765).
- The method outperforms state-of-the-art approaches on both real and synthetic datasets while maintaining computational feasibility on standard CPUs.
Transmittance-Guided Structure-Texture Decomposition for Nighttime Image Dehazing
Introduction
Nighttime hazy image degradation is a multifactorial phenomenon, encompassing atmospheric scattering, absorption by suspended aerosols, artificial lighting-induced non-uniform illumination, and substantial chromatic distortions. Unlike daytime dehazing, where atmospheric light and haze distribution are relatively uniform and well-modeled, nighttime scenarios require algorithms to handle spatially variant illumination and complex glow effects from artificial sources. Prior approaches, including both traditional and deep learning methods, have improved visibility and suppressed glow or color bias partially, but most fail to jointly resolve haze removal, brightness normalization, color restoration, and detail enhancement.
Methodology
Two-Stage Dehazing Pipeline
The proposed framework consists of a two-stage pipeline integrating physically-grounded transmittance correction and structure-texture layered optimization:
Stage 1: Transmittance-Based Dehazing
Initial transmittance maps are obtained using boundary constraint methods; these are corrected via region-adaptive compensation (for bright regions and artificial light sources) and normalization. Atmospheric light distribution is estimated using a quadratic Gaussian filter in the YUV color space (leveraging the gray haze-line prior for improved haze mapping) and combined with the corrected transmittance map in an improved nighttime imaging model that allows spatially varying atmospheric light.
Stage 2: Structure-Texture Optimization (STAR-YUV Decomposition)
The dehazed output is decomposed into structure and texture layers using an adaptation of the STAR model in the YUV domain. The structure layer is enhanced via gamma correction (for illumination compensation) and MSRCR (Multi-Scale Retinex with Color Restoration) for chromatic bias correction, while the texture layer undergoes Laplacian-of-Gaussian filtering to reinforce detail and edge features.
Fusion Strategy
Two-phase fusion combines enhancements: nonlinear Retinex-based fusion (multiplicative combination) of structure and texture, followed by linear blending with the initial dehazed result. This balances brightness, color fidelity, and detail preservation—mitigating over-brightening, whitening artifacts, and haze residue.
Experimental Evaluation
Datasets and Metrics
Experiments leverage the ZS330 (real-world) and HC770 (synthetic) nighttime hazy image datasets. Benchmark methods span both daytime (e.g., DCP [he2011single], DehazeNet [cai2016dehazenet]) and nighttime (e.g., GMLC [li2015nighttime], CEEF [liu2022joint], IAT [cui2022illumination], Fb [zhou2023lowlight]) dehazing. Evaluation uses full-reference (PSNR, SSIM) and no-reference (AG, IE, NIQE) metrics, providing assessment of pixel-level fidelity, structural quality, perceptual naturalness, and detail richness.
Color Restoration
Controlled experiments using CIEDE2000 color charts demonstrate a significant improvement in color fidelity. The method achieves an average color difference of 15.466, a 20.1% improvement over the next best competitor, confirming effective chromatic restoration beyond previous algorithms which either oversaturate, bias colors, or lose fidelity under low illumination.
Subjective and Objective Comparisons
Subjective evaluation by 30 assessors shows highest scores in all categories: detail (4.0), color (4.3), naturalness (4.0), visual appeal (4.2), overall (4.1; on a 1–5 scale), substantiating the algorithm's visual effectiveness.
Quantitative results:
- PSNR: 17.024 dB (highest vs. competitors; indicates closest pixel-wise resemblance to haze-free ground truth)
- SSIM: 0.765 (highest structural similarity)
- AG/IE: 7.604/7.528 (best gradient/detail richness and information content)
- NIQE: 2.693 (second-best perceptual quality; consistently strong across all metrics)
Experimental results are robust across both synthetic and real-world datasets, outperforming state-of-the-art daytime and nighttime methods in both color recovery and detail preservation.
Ablation Study
Component-wise removal confirms the necessity and complementarity of each stage: transmittance correction (w/o T) increases noise, structure-texture optimization (w/o STAR) leaves unresolved chromatic and illumination artifacts, and omitting dehazing (w/o Dehaze) leads to over-brightened yet still hazy outputs. Only the full pipeline meets all quality criteria.
Computational Efficiency
Runtime analysis indicates comparable speed to complex baselines (e.g., GMLC, IAT), though not as fast as lightweight approaches (CEEF, DehazeNet), with substantial gains in output quality justifying the additional computation. The pipeline is implementable on standard CPUs, with future optimization avenues identified.
Implications and Future Directions
The framework establishes a principled, physically-grounded approach that integrates region-aware transmittance correction and structure-texture disentanglement in the YUV domain. Results:
- Jointly resolve glow suppression, haze removal, color restoration, and detail enhancement;
- Avoid reliance on massive paired datasets typical in deep learning approaches, while preserving interpretability via explicit physical priors;
- Set new benchmarks for nighttime dehazing in both objective and subjective metrics.
Potential future developments include:
- Algorithmic acceleration and hardware-aware optimizations for real-time processing;
- Extension to temporal (video) dehazing, leveraging spatial-temporal consistency;
- Incorporation into higher-level vision frameworks (e.g., pedestrian detection, surveillance, scene parsing) for robust downstream analysis under adverse nighttime conditions;
- Hybridization with deep learning models exploiting both physical priors and learned representations for further performance and generalization gains.
Conclusion
This work introduces a comprehensive two-stage nighttime image dehazing method that integrates region-adaptive transmittance correction with STAR-YUV structure-texture decomposition and targeted enhancement. Demonstrated superior performance in color fidelity, structural preservation, and detail richness, as indicated by strong numerical metrics and subjective scores. Future research should focus on real-time optimization and broader integration with multimodal visual perception systems.
Reference: "Transmittance-Guided Structure-Texture Decomposition for Nighttime Image Dehazing" (2603.29507).