Abstract
Single-image dehazing is a challenging task because haze density typically increases exponentially with scene depth, leading to non-uniform degradation across a single image. Traditional multi-stage deep learning models often rely on a "coarse-to-fine" strategy that applies uniform supervision across all stages, which can fail to capture these varying levels of haze. This paper introduces LAP-Net (Level-Aware Progressive Network), a novel architecture that handles gradually increasing levels of haze through a multi-stage network. By employing a progressive transmission subnet and an adaptive integration scheme guided by local entropy and haze-level classification, LAP-Net achieves superior restoration of both foreground details and heavily obscured backgrounds. Experimental results on the RESIDE and O-HAZE datasets demonstrate state-of-the-art performance.
Introduction
Presented at ICCV 2019, LAP-Net moves away from traditional iterative sharpening methods. Instead, it proposes a level-aware approach where different stages of the network are explicitly trained to handle specific haze intensities, ranging from mild to dense. This allows the model to effectively restore clear details in both nearby objects and distant, heavily obscured backgrounds.
The Challenge of Non-Uniform Haze
Image dehazing is inherently difficult because haze is rarely uniform. According to the atmospheric scattering model, haze density increases with scene depth. In a single photograph, the foreground might exhibit "mild" haze, while the background is buried under "dense" haze.
Most existing multi-stage models use a coarse-to-fine strategy: every stage tries to sharpen the previous stage's output using the final clear image as supervision. The authors argue this is suboptimal because it doesn't account for varying degradation levels; a stage optimized for mild haze may fail on distant buildings, while one optimized for dense haze might over-process the foreground.
Methodology: The LAP-Net Architecture
LAP-Net consists of four main components integrated into an end-to-end pipeline:
- Progressive Transmission Subnet (t-net): Instead of one-size-fits-all training, the t-net uses cascaded "hourglass" units. Each stage is supervised by a transmission map corresponding to a specific haze level (from mild to dense), allowing the network to learn the "easy" haze before tackling the "hard" haze.
- Discrete Atmospheric Light Subnet (A-net): Estimating global atmospheric light ($A$) is treated as a classification task rather than regression. By discretizing possible values into 100 bins, the model achieves a more precise and robust estimation, avoiding "regression-to-mean" errors.
- Restoration Layer: This layer takes the transmission maps from every stage of the t-net and the atmospheric light from the A-net to produce multiple "candidate" dehazed images—one for each stage.
- Adaptive Integration Subnet (I-net): This component fuses the candidate images using local entropy guidance (identifying regions with the best-restored textures), a memory network for spatial smoothness, and a contextual weighting mechanism that considers the overall haze level of the input.
Key Results and Insights
LAP-Net was evaluated using several key datasets, including the NYU Depth Dataset v2 for synthesis, RESIDE (SOTS-outdoor) for standard benchmarks, and O-HAZE for real-world dense haze. The model achieved state-of-the-art performance:
- RESIDE: Achieved an SSIM of 0.934 and a CIEDE2000 of 5.864, outperforming popular models like AOD-Net.
- O-HAZE: On this tougher dataset, it achieved an SSIM of 0.798, a significant jump over previous methods like PDN (0.739).
- Qualitative Quality: Visually, LAP-Net avoids common pitfalls such as over-saturation or over-exposure in sky regions, maintaining high color fidelity even in distant, hazy areas.
Conclusion
LAP-Net demonstrates that level-aware progressive restoration is more effective than simple iterative refinement for non-uniform degradation. By incorporating domain knowledge—such as local entropy—into the attention mechanism and treating global parameter estimation as a classification problem, the network provides a robust solution for single-image dehazing. Future work may explore reducing computational complexity for real-time mobile deployment and extending this hierarchical integration scheme to other low-level vision tasks like deraining or denoising.
Summary updated