Abstract
Due to the lack of natural scene and haze prior information, it is greatly challenging to completely remove the haze from a single image without distorting its visual content. Fortunately, the real-world haze usually presents non-homogeneous distribution, which provides us with many valuable clues in partial well-preserved regions. In this paper, we propose a Non-Homogeneous Haze Removal Network (NHRN) via artificial scene prior and bidimensional graph reasoning. Firstly, we employ the gamma correction iteratively to simulate artificial multiple shots under different exposure conditions, whose haze degrees are different and enrich the underlying scene prior. Secondly, beyond utilizing the local neighboring relationship, we build a bidimensional graph reasoning module to conduct non-local filtering in the spatial and channel dimensions of feature maps, which models their long-range dependency and propagates the natural scene prior between the well-preserved nodes and the nodes contaminated by haze. To the best of our knowledge, this is the first exploration to remove non-homogeneous haze via the graph reasoning based framework. We evaluate our method on different benchmark datasets. The results demonstrate that our method achieves superior performance over many state-of-the-art algorithms for both the single image dehazing and hazy image understanding tasks.
Introduction
Single image dehazing is a fundamental computer vision task that aims to restore a clear scene from a hazy input. Most traditional algorithms rely on the atmospheric scattering model, which assumes haze is uniformly distributed. However, real-world haze is often non-homogeneous—patchy and varying in density across the image. Standard models often fail in these scenarios, either leaving artifacts in thick haze regions or distorting clear areas. The proposed Non-Homogeneous Haze Removal Network (NHRN) addresses these limitations by leveraging artificial scene priors and non-local contextual information through graph reasoning.
Methodology
The NHRN architecture is designed to overcome the lack of depth information and scene priors in single-image dehazing through two primary components:
Artificial Scene Prior (AMS Module)
To simulate the benefit of having multiple exposures of the same scene, the Artificial Multiple Shots (AMS) module uses iterative gamma correction on the single input image. This process generates a sequence of images with varying exposure levels, effectively revealing structural details in hazy regions that were previously obscured. These shots are then integrated using a Gated Fusion (GF) module, which adaptively weights the most informative features from each simulated exposure to enrich the underlying scene prior.
Bidimensional Graph Reasoning
Standard convolutional neural networks (CNNs) are limited by their local receptive fields. NHRN introduces a bidimensional graph reasoning module to capture long-range dependencies:
- Spatial Graph Reasoning (SGR): This module groups similar pixels into nodes regardless of their spatial distance. It allows the network to propagate information from clear regions to hazy ones that share similar semantic content (e.g., using a clear patch of grass to help reconstruct a hazy one).
- Channel Graph Reasoning (CGR): This component models dependencies between different feature channels, capturing global content-related information such as edges and textures.
Experimental Results
The NHRN was evaluated on several benchmarks, including the NH-HAZE dataset for non-homogeneous haze and the RESIDE dataset for synthetic haze.
Quantitative Performance
On the SOTS (Synthetic Object Test Set), NHRN achieved a significant performance boost, with a 2.46dB gain in PSNR and a 0.0022 gain in SSIM over previous state-of-the-art methods. Furthermore, the model's utility was tested on downstream tasks like object detection in foggy conditions (using Foggy Cityscapes-DBF), where it achieved the highest Mean Average Precision (mAP) across various haze concentrations.
Qualitative Comparison
Visual results demonstrate that NHRN effectively removes non-homogeneous haze while preserving natural colors and textures. Unlike scattering-model-based methods (e.g., DCPDN) or standard end-to-end models (e.g., GridNet), NHRN avoids common artifacts like color distortion or residual haze patches.
Conclusion
NHRN represents a novel approach to non-homogeneous haze removal by combining artificial exposure simulation with non-local graph reasoning. By "borrowing" details from clear regions to restore hazy ones, the network achieves superior restoration quality. Future work will focus on reducing computational redundancy and extending the framework to video dehazing by exploiting temporal correlations.
Summary updated