Abstract
Real-world image dehazing remains a significant challenge in computer vision due to the inherent domain gap between synthetic training data and the complex physics of natural haze. This paper presents WTCL-Dehaze, a semi-supervised framework that integrates Discrete Wavelet Transforms (DWT) and Contrastive Learning to bridge this gap. By utilizing a dual-branch architecture that learns from both labeled synthetic pairs and unlabeled real-world images, the model preserves high-frequency textures and global structures more effectively than previous methods. Experimental results on the RESIDE dataset demonstrate that WTCL-Dehaze achieves state-of-the-art performance, reaching a PSNR of 27.24 on the SOTS outdoor benchmark.
Introduction
Images captured in hazy outdoor environments often suffer from severe degradation, including color distortion, reduced contrast, and a loss of fine-grained detail. In computer vision, the task of "dehazing" aims to recover the original scene radiance from these degraded inputs. However, dehazing is a fundamentally ill-posed problem. According to the standard atmospheric scattering model, a hazy image is a composite of the clear scene, a transmission map, and global atmospheric light. With only the hazy image as input, an infinite number of potential solutions exist.
While deep learning has advanced the field, most existing models are trained on synthetic datasets where haze is mathematically simulated. These models frequently struggle when applied to real-world scenarios because the simulated haze fails to capture the complex, non-uniform nature of actual atmospheric conditions—a phenomenon known as the domain gap. WTCL-Dehaze addresses this by rethinking the dehazing pipeline through a semi-supervised lens, focusing on frequency-aware feature extraction and contrastive regularization.
The WTCL-Dehaze Framework
WTCL-Dehaze employs a dual-branch semi-supervised learning strategy designed to maximize the utility of available data while ensuring generalization to real-world environments.
1. The Supervised Branch
This branch is trained using labeled synthetic pairs (hazy images and their corresponding ground truths). It utilizes standard reconstruction losses, such as L1 and Mean Squared Error (MSE), alongside a VGG-19 based perceptual loss to ensure the restored images are both pixel-accurate and visually consistent with human perception.
2. The Unsupervised Branch
To tackle the domain gap, the unsupervised branch processes unlabeled real-world hazy images. Since no ground truth is available for these images, the model relies on physical "priors" to guide the restoration process. Specifically, it incorporates Dark Channel Loss and Total Variation (TV) loss, which enforce the characteristics of natural, clear images—such as the presence of dark pixels in local patches and spatial smoothness—onto the output.
Methodological Innovations
Two core components distinguish WTCL-Dehaze from traditional convolutional neural networks (CNNs):
Discrete Wavelet Transform (DWT)
Standard CNN architectures often lose critical high-frequency information during downsampling operations like pooling or strided convolutions. WTCL-Dehaze replaces these operations with the Discrete Wavelet Transform. DWT decomposes feature maps into high-frequency components (capturing edges and textures) and low-frequency components (capturing global scene structure). This allows the network to process fine details and overall layout separately before reconstructing the final image via Inverse DWT (IDWT), resulting in significantly sharper edges.
Contrastive Regularization (CR)
Contrastive learning is utilized as a powerful regularizer to improve feature representation. In this framework, the restored image is treated as an "anchor." The model is trained to pull this anchor closer to the clear "positive" image while pushing it further away from the hazy "negative" image in a high-dimensional latent space. This comparison-based learning helps the model better distinguish between hazy artifacts and true scene features, leading to more natural color reproduction.
Experimental Results
The model was evaluated using the RESIDE (Realistic Single Image Dehazing) dataset. Training involved 4,000 labeled images from the Indoor (ITS) and Outdoor (OTS) sets, combined with 2,000 unannotated real hazy images (URHI). Benchmarking was performed on the Synthetic Object Testing Set (SOTS) and the Hybrid Subjective Testing Set (HSTS).
WTCL-Dehaze outperformed 11 leading dehazing methods, including traditional approaches like the Dark Channel Prior (DCP) and deep learning models like AOD-Net and CycleGAN. On the SOTS outdoor dataset, the model achieved a PSNR of 27.24 and an SSIM of 0.971. This represents a 1.97 dB improvement in PSNR over the previous leading semi-supervised method.
Qualitatively, WTCL-Dehaze produces images with superior clarity and fewer artifacts. Unlike GAN-based methods, which often introduce "hallucinated" textures or unnatural color shifts, WTCL-Dehaze maintains high fidelity to the original scene's colors and structures.
Conclusion and Future Work
WTCL-Dehaze demonstrates that combining frequency-domain analysis with semi-supervised contrastive learning effectively mitigates the domain gap in image dehazing. By preserving high-frequency details through DWT and refining feature spaces through contrastive regularization, the model provides a robust solution for real-world visibility enhancement.
Future research directions include the integration of Vision Transformers (ViTs) to better capture long-range global context and the adaptation of this framework to other low-visibility restoration tasks, such as underwater imaging and low-light enhancement. Additionally, optimizing the computational efficiency of the multi-loss training regime remains a priority for deploying these models on edge devices.
Summary updated