[Paper Review] Single Image Haze Removal using a Generative Adversarial Network

N. Bharath Raj, N. Venkateswaran|arXiv (Cornell University)|Oct 22, 2018

Image Enhancement Techniques25 references20 citations

TL;DR

This paper proposes an end-to-end generative adversarial network (GAN)-based method for single-image haze removal, using a modified Tiramisu network as the generator and a patch-based discriminator to reduce artifacts. A hybrid weighted loss function enhances perceptual quality, achieving state-of-the-art performance on both synthetic and real-world hazy images.

ABSTRACT

Traditional methods to remove haze from images rely on estimating a transmission map. When dealing with single images, this becomes an ill-posed problem due to the lack of depth information. In this paper, we propose an end-to-end learning based approach which uses a modified conditional Generative Adversarial Network to directly remove haze from an image. We employ the usage of the Tiramisu model in place of the classic U-Net model as the generator owing to its higher parameter efficiency and performance. Moreover, a patch based discriminator was used to reduce artefacts in the output. To further improve the perceptual quality of the output, a hybrid weighted loss function was designed and used to train the model. Experiments on synthetic and real world hazy images demonstrates that our model performs competitively with the state of the art methods.

Motivation & Objective

• To address the ill-posed nature of single-image haze removal due to missing depth information.
• To develop an end-to-end learning-based approach that bypasses transmission map estimation.
• To improve perceptual quality and reduce artifacts in dehazed outputs.
• To achieve competitive performance on both synthetic and real-world hazy images.
• To leverage the parameter efficiency and performance of the Tiramisu architecture in a GAN framework for dehazing.

Proposed method

• The generator is based on the Tiramisu network, chosen for its high parameter efficiency and strong feature learning capability.
• A patch-based discriminator is employed to enforce local realism and reduce texture artifacts in the dehazed output.
• A hybrid weighted loss function combines content, adversarial, and perceptual losses to improve visual quality.
• The model is trained end-to-end in a conditional GAN setting, directly mapping hazy images to dehazed ones.
• The Tiramisu generator replaces the standard U-Net to improve performance and reduce parameters.
• Training uses a combination of L1 loss, adversarial loss, and perceptual loss to balance fidelity and realism.

Experimental results

Research questions

• RQ1 Can a GAN-based approach effectively remove haze from single images without requiring depth information?
• RQ2 Does using the Tiramisu network as a generator improve performance and efficiency compared to U-Net in dehazing tasks?
• RQ3 Can a patch-based discriminator reduce artifacts and improve local image quality in dehazed outputs?
• RQ4 To what extent does a hybrid loss function enhance perceptual quality compared to standard loss functions?
• RQ5 How does the proposed method compare to state-of-the-art methods on real-world and synthetic hazy images?

Key findings

• The proposed method achieves competitive performance on both synthetic and real-world hazy image datasets.
• The use of the Tiramisu generator leads to better parameter efficiency and improved feature representation compared to U-Net.
• The patch-based discriminator significantly reduces texture artifacts and improves local image quality.
• The hybrid weighted loss function enhances perceptual quality, resulting in more natural-looking dehazed images.
• The model demonstrates strong generalization on real-world hazy images, outperforming several existing methods in visual quality.
• The method is validated and accepted for presentation at WiSPNET 2020, with ongoing updates available on the associated GitHub repository.