Dense '123' Color Enhancement Dehazing Network

Tiantong Guo, Venkateswararao Cherukuri, Vishal MongaWorkshops @ CVPR 2019 · DOI

Introduction

Single image dehazing is a fundamental computer vision task aimed at recovering a clear scene from an image obscured by atmospheric particles. Traditional methods often rely on the classical haze model, which decomposes an image into scene radiance, a transmission map, and global atmospheric light. However, these models frequently fail when confronted with extremely dense haze, such as heavy fog or smoke. In these scenarios, color information is severely degraded, and standard single-path encoder-decoder networks often produce "washed out" or desaturated results.

Comparison of a clear scene and a dense haze environment, highlighting the extreme loss of color and structural detail.

To address these limitations, the Dense ‘123’ Color Enhancement Dehazing Network (123-CEDH) was developed. This architecture explicitly prioritizes color recovery through a unique multi-decoder design, enabling it to secure a Top-3 position in the NTIRE 2019 Dehazing Challenge.

The "One-to-Three" Architecture

The core innovation of 123-CEDH is its structural flow: one shared encoder branching into three independent decoders.

  1. Shared Encoder: The network utilizes a DenseNet-121 backbone to extract high-level features from the hazy input.
  2. Independent Color Decoders: Instead of a single bottleneck, the features are passed to three distinct DenseNet-based decoders. Each decoder is dedicated to estimating a single color channel: Red, Green, or Blue. This separation allows the network to learn specific scattering patterns associated with different light wavelengths.
  3. Refinement Block: The outputs from the three decoders are concatenated and passed through a final refinement convolutional block. This stage merges the channel-specific information, eliminates potential artifacts, and ensures global color cohesion.
The 123-CEDH network architecture, showing the transition from a single DenseNet-121 encoder to three independent color channel decoders.

Compound Loss and Training Strategy

The network is optimized using a sophisticated four-part compound loss function designed to balance pixel accuracy with perceptual realism:

  • Reconstruction Loss ($\ell_2$): Maintains standard pixel-wise fidelity.
  • Perceptual Loss (VGG): Ensures the recovered textures and content are visually realistic.
  • Multi-Scale Structural Similarity (MS-SSIM): Preserves structural integrity across various image scales.
  • Contrast Enhancement Loss ($L_{ce}$): A novel term that maximizes the variance within individual color channels. This prevents the "flat" or gray appearance common in dehazed images by incentivizing the network to recover the full dynamic range of colors.

Training is conducted in two stages. First, the encoder is pre-trained with a single decoder to establish a baseline. Subsequently, the full three-decoder architecture is trained to fine-tune color-specific recovery.

Key Insights and Results

The 123-CEDH approach yielded several critical insights:

  • Early Color Recovery: The split-channel architecture begins recovering accurate color information much earlier in the training process than standard models, which typically prioritize luminance first.
  • Variance as Clarity: The Contrast Enhancement Loss effectively forces the model to avoid "graying out" the scene, resulting in images with significantly higher saturation and "pop."

In the NTIRE 2019 Challenge, 123-CEDH achieved a PSNR of 17.10 dB on the validation set, a substantial improvement over previous state-of-the-art methods that typically reached only 11-13 dB on dense haze tasks. Ablation studies confirmed that the move from a single decoder to three independent decoders provided a performance boost of approximately 0.75 dB.

Conclusion

The 123-CEDH network demonstrates that explicit channel-wise modeling is a powerful strategy for overcoming the extreme degradation caused by dense haze. While the three-decoder structure increases computational overhead, the significant gains in color accuracy and structural clarity make it a robust solution for challenging atmospheric conditions. Future work may focus on optimizing the model for mobile deployment and exploring its efficacy on other distortions like rain or snow.

Summary updated

BibTeX
@inproceedings{guo2019dense,
  author = {Tiantong Guo and Venkateswararao Cherukuri and Vishal Monga},
  title = {{Dense '123' Color Enhancement Dehazing Network}},
  booktitle = {Workshops @ CVPR},
  year = {2019},
  doi = {10.1109/cvprw.2019.00266},
}
Figure