Perceptually Optimized Generative Adversarial Network for Single Image Dehazing

Yixin Du, Xin Li

Abstract: Existing approaches towards single image dehazing including both model-based and learning-based heavily rely on the estimation of so-called transmission maps. Despite its conceptual simplicity, using transmission maps as an intermediate step often makes it more difficult to optimize the perceptual quality of reconstructed images. To overcome this weakness, we propose a direct deep learning approach toward image dehazing bypassing the step of transmission map estimation and facilitating end-to-end perceptual optimization. Our technical contributions are mainly three-fold. First, based on the analogy between dehazing and denoising, we propose to directly learn a nonlinear mapping from the space of degraded images to that of haze-free ones via recursive deep residual learning; Second, inspired by the success of generative adversarial networks (GAN), we propose to optimize the perceptual quality of dehazed images by introducing a discriminator and a loss function adaptive to hazy conditions; Third, we propose to remove notorious halo-like artifacts at large scene depth discontinuities by a novel application of guided filtering. Extensive experimental results have shown that the subjective qualities of dehazed images by the proposed perceptually optimized GAN (POGAN) are often more favorable than those by existing state-of-the-art approaches especially when hazy condition varies.

Introduction

Single image dehazing is a critical task in computer vision, traditionally solved using physical models that estimate a "transmission map" to describe light scattering. However, this approach suffers from three major flaws: model mismatch in complex lighting, error propagation from map estimation to final reconstruction, and a perceptual gap where mathematical optimization doesn't align with human visual preference.

To address these, we introduce POGAN (Perceptually Optimized Generative Adversarial Network), a framework that treats haze as structured noise and optimizes for visual realism directly, bypassing intermediate transmission maps entirely.

The POGAN Architecture

The core of POGAN is a recursive deep residual learning network. By reformulating the standard haze equation, we treat dehazing as a denoising problem. The generator uses 16 recursive sub-blocks to progressively subtract haze from the input image. This architecture allows the network to learn a direct mapping from hazy to clear space without the constraints of simplified physical models.

The POGAN generator architecture utilizing recursive residual blocks for progressive haze subtraction.

Perceptual Optimization and Halo Removal

To ensure the output looks natural to human observers, POGAN employs a Generative Adversarial Network (GAN) setup. A discriminator network evaluates the realism of the dehazed images, while an adaptive perceptual loss function adjusts its weights based on haze severity. In thick haze, the model prioritizes aggressive removal; in light haze, it focuses on preserving fine textures.

Furthermore, deep learning models often produce "halo" artifacts at sharp depth edges. We mitigate this by applying a guided filter to the residual map—the difference between the hazy input and dehazed output—to preserve edges and smooth out artifacts.

Ablation study demonstrating the visual improvements gained by adding the Discriminator (D) and Halo removal (H) modules to the base Generator (G).

Experimental Results

We evaluated POGAN using a mix of synthetic data (derived from DIV2K and Middlebury datasets) and real-world "in the wild" images.

  • Subjective Quality: POGAN consistently outperforms state-of-the-art methods like Dark Channel Prior (DCP) and AOD-Net in visual comparisons, producing images with superior color fidelity and fewer artifacts in foreground regions.
  • Robustness: The adaptive loss function allows the model to handle varying haze thicknesses effectively, avoiding the common pitfalls of over-dehazing or under-dehazing found in traditional models.
  • Quantitative Performance: While GAN-based methods are sometimes penalized by traditional metrics like PSNR due to the textures they generate, POGAN remains highly competitive while delivering significantly better subjective visual results.

Conclusion

By removing the reliance on intermediate transmission maps, POGAN simplifies the dehazing pipeline and enables direct perceptual optimization. The combination of recursive residual learning, adaptive GAN loss, and guided filtering results in a robust system capable of producing vivid, natural-looking images across a wide range of hazy conditions. Future work will focus on training with paired real-world data to further reduce reliance on synthetic haze models.

Summary updated

Abstract

Existing approaches towards single image dehazing including both model-based and learning-based heavily rely on the estimation of so-called transmission maps. Despite its conceptual simplicity, using transmission maps as an intermediate step often makes it more difficult to optimize the perceptual quality of reconstructed images. To overcome this weakness, we propose a direct deep learning approach toward image dehazing bypassing the step of transmission map estimation and facilitating end-to-end perceptual optimization. Our technical contributions are mainly three-fold. First, based on the analogy between dehazing and denoising, we propose to directly learn a nonlinear mapping from the space of degraded images to that of haze-free ones via recursive deep residual learning; Second, inspired by the success of generative adversarial networks (GAN), we propose to optimize the perceptual quality of dehazed images by introducing a discriminator and a loss function adaptive to hazy conditions; Third, we propose to remove notorious halo-like artifacts at large scene depth discontinuities by a novel application of guided filtering. Extensive experimental results have shown that the subjective qualities of dehazed images by the proposed perceptually optimized GAN (POGAN) are often more favorable than those by existing state-of-the-art approaches especially when hazy condition varies.
BibTeX
@misc{du2018perceptually,
  author = {Yixin Du and Xin Li},
  title = {{Perceptually Optimized Generative Adversarial Network for Single Image Dehazing}},
  year = {2018},
  eprint = {1805.01084},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1805.01084},
}
Figure