UR2P-Dehaze: Learning a Simple Image Dehaze Enhancer via Unpaired Rich Physical Prior

TL;DR

UR2P-Dehaze tackles single-image dehazing without paired data by learning rich priors directly from hazy images. It integrates a Retinex-guided Shared Prior Estimator, Dynamic Wavelet Separable Convolution for multi-scale feature fusion in the wavelet domain, and an Adaptive Color Corrector to restore authentic colors, all guided by a composite loss that stabilizes decomposition and reconstruction. The approach yields state-of-the-art quantitative results across multiple datasets and improves downstream tasks like object detection, while maintaining a concise and generalizable architecture. This work demonstrates that unpaired, physics-inspired priors combined with wavelet-domain processing can robustly restore hazy images under diverse conditions with practical deployment potential.

Abstract

Image dehazing techniques aim to enhance contrast and restore details, which are essential for preserving visual information and improving image processing accuracy. Existing methods rely on a single manual prior, which cannot effectively reveal image details. To overcome this limitation, we propose an unpaired image dehazing network, called the Simple Image Dehaze Enhancer via Unpaired Rich Physical Prior (UR2P-Dehaze). First, to accurately estimate the illumination, reflectance, and color information of the hazy image, we design a shared prior estimator (SPE) that is iteratively trained to ensure the consistency of illumination and reflectance, generating clear, high-quality images. Additionally, a self-monitoring mechanism is introduced to eliminate undesirable features, providing reliable priors for image reconstruction. Next, we propose Dynamic Wavelet Separable Convolution (DWSC), which effectively integrates key features across both low and high frequencies, significantly enhancing the preservation of image details and ensuring global consistency. Finally, to effectively restore the color information of the image, we propose an Adaptive Color Corrector that addresses the problem of unclear colors. The PSNR, SSIM, LPIPS, FID and CIEDE2000 metrics on the benchmark dataset show that our method achieves state-of-the-art performance. It also contributes to the performance improvement of downstream tasks. The project code will be available at https://github.com/Fan-pixel/UR2P-Dehaze. \end{abstract}

Figures (13)

• Figure 1: RGB three-channel comparison of the image with haze and the image with the haze removed. (a) Represents the visual comparison between the hazy image, the dehaze image, and the clear image. (b) Representation of the difference in the RGB three channels between the hazy image and the clear image. (c) Difference between the dehaze image and the clear image in the RGB three channels.
• Figure 2: The average PSNR and SSIM are used as evaluation metrics to compare the performance of the proposed unsupervised dehazing method with other state-of-the-art methods on the SOTS-outdoor dataset.
• Figure 3: The overall pipeline of our UR2P-Dehaze. It comprises (a) an adaptive learning branch that estimates illumination, reflectance, and color priors, (b) a cascaded frequency decomposition process using wavelet transform, and (c) The dynamic wavelet separable convolution (DWSC) module leverages physical features to dynamically model the image, adaptively combining frequency and spatial domain information to refine dehazing results and enhance realism and color fidelity. Additionally, the feature refinement network incorporates both a residual module and a feature fusion module to further enhance feature representation and integration.
• Figure 4: The process oof performing convolution operation in wavelet domain. It can effectively reduce the number of model parameters and improve the multi-scale feature extraction ability of the model.
• Figure 5: Visual comparison of haze removal on samples from the SOTS-indoor dataset.