Abstract:<p>Single image dehazing is an important problem as the existence of haze degrades the quality of the image and hinders most high-level computer vision tasks. Previous methods solve this problem using various low-level statistics priors or learning on synthetic data sets with CNN. In practice, the low-level priors are not always held in various scenes. And many CNN based methods directly estimate the transmission maps and atmospheric lights from huge synthetic data. However, without the guidance or constraints of priors may lead to over-dehazed or under-dehazed results. To address these issues, we propose a prior guided conditional generative adversarial network, an end-to-end model that generates realistic clean images using hazy image input and dehazed image based on the traditional prior-based method. The proposed generator extracted the feature with a parameters-shared encoder, and the clear image is recovered by decoding multi-scale features, which are fused and enhanced by the proposed attention-based feature aggregation block. And two-scale discriminators are adopted to supervise the generator to recover more image details with a combination of perceptual loss and adversarial loss. Our algorithm can efficiently combine the prior-based and CNN based image dehazing method and remove the weakness of each other. Experimental results on synthetic datasets and real-world images demonstrate our model can generate more perceptually appealing dehazing results, and provide superior performance compared with the state-of-the-art methods.</p>

What problem does this paper attempt to address?

### Main Problem Addressed by the Paper This paper proposes a new solution to the problem of single image dehazing. Specifically, the authors introduce a Prior Guided Conditional Generative Adversarial Network (PG-CGAN) to address the limitations of traditional dehazing methods and the shortcomings of deep learning-based methods when dealing with real-world images. ### Summary of Core Contributions 1. **Proposing the PG-CGAN Framework**: Utilizing the dehazing results of the dark channel prior as guiding information, extracting useful features through an attention mechanism, and employing an adaptive group normalization module to improve partial correct style transfer during the decoding phase, thereby effectively enhancing the visual quality of the restored image. 2. **Developing a Multi-Scale Discriminator**: To supervise the generator in recovering more details, a dual-scale discriminator is designed, and a random cropping method is used to judge the authenticity of the generated image, making the generator more focused on local changes. 3. **Comprehensive Loss Function**: Combining structural similarity loss, pixel-level loss, and adversarial loss to construct the objective function, optimizing the network to produce more perceptually appealing dehazed images. 4. **Extensive Experimental Validation**: Experiments are conducted on various datasets, including synthetic data and real-world images, demonstrating the effectiveness and superiority of the proposed method. ### Methodological Background The paper first reviews the research progress of Generative Adversarial Networks (GANs) and then discusses existing single image dehazing methods, categorizing them into model-based methods and learning-based methods. Model-based methods rely on low-level statistical prior knowledge, while learning-based methods focus on estimating parameters from training data. The paper points out that although these methods have their advantages, they also have certain limitations, such as prior knowledge not being applicable to all scenarios or performance degradation when dealing with real-world images. ### Technical Solution - **Encoder-Decoder Architecture**: Using a parameter-sharing encoder to extract features from the input image (hazy image and dehazing result based on the dark channel prior), and then restoring the clear image through the decoder. - **Attention Mechanism**: Introducing an attention mechanism to select useful information from the guiding image, reducing the impact of noise and artifacts. - **Adaptive Group Normalization**: Employing adaptive group normalization during the decoding process to assist in partial correct style transfer. - **Multi-Scale Discriminator**: Supervising the generator through a multi-scale discriminator to ensure the realism and detail richness of the generated image. In summary, the method proposed in this paper aims to combine the advantages of traditional prior knowledge with the powerful representation capabilities of deep learning to overcome the limitations of existing dehazing methods and perform excellently in practical applications.

Prior guided conditional generative adversarial network for single image dehazing

From Coarse to Fine: A Two-Stage Network for Dense Haze removal

Unsupervised single image dehazing with generative adversarial network

Remote sensing image dehazing using generative adversarial network with texture and color space enhancement

Haze Relevant Feature Attention Network for Single Image Dehazing

Enhanced CycleGAN Network with Adaptive Dark Channel Prior for Unpaired Single-Image Dehazing

FD-GAN: Generative Adversarial Networks with Fusion-discriminator for Single Image Dehazing

Dehaze-GLCGAN: Unpaired Single Image De-hazing via Adversarial Training

Image dehazing algorithm based on improved generative adversarial network

Residual Spatial and Channel Attention Networks for Single Image Dehazing

Single image dehazing via cycle-consistent adversarial networks with a multi-scale hybrid encoder-decoder and global correlation loss

Perceptually Optimized Generative Adversarial Network for Single Image Dehazing

GGADN: Guided generative adversarial dehazing network

Single image dehazing using generative adversarial networks based on an attention mechanism

SIDGAN: Single Image Dehazing Without Paired Supervision.

Physical-model guided self-distillation network for single image dehazing

Single Image Dehazing Based on Two-Stream Convolutional Neural Network

PriorNet: A Novel Lightweight Network with Multidimensional Interactive Attention for Efficient Image Dehazing

A robust edge-guided and color-guided network for single image dehazing

Single image dehazing algorithm based on generative adversarial network

DFC-dehaze: an Improved Cycle-Consistent Generative Adversarial Network for Unpaired Image Dehazing