Abstract:Image dehazing has drawn a significant attention in recent years. Learning-based methods usually require paired hazy and corresponding ground truth (haze-free) images for training. However, it is difficult to collect real-world image pairs, which prevents developments of existing methods. Although several works partially alleviate this issue by using synthetic datasets or small-scale real datasets. The haze intensity distribution bias and scene homogeneity in existing datasets limit the generalization ability of these methods, particularly when encountering images with previously unseen haze intensities. In this work, we present LMHaze, a large-scale, high-quality real-world dataset. LMHaze comprises paired hazy and haze-free images captured in diverse indoor and outdoor environments, spanning multiple scenarios and haze intensities. It contains over 5K high-resolution image pairs, surpassing the size of the biggest existing real-world dehazing dataset by over 25 times. Meanwhile, to better handle images with different haze intensities, we propose a mixture-of-experts model based on Mamba (MoE-Mamba) for dehazing, which dynamically adjusts the model parameters according to the haze intensity. Moreover, with our proposed dataset, we conduct a new large multimodal model (LMM)-based benchmark study to simulate human perception for evaluating dehazed images. Experiments demonstrate that LMHaze dataset improves the dehazing performance in real scenarios and our dehazing method provides better results compared to state-of-the-art methods.

What problem does this paper attempt to address?

The paper attempts to address several key issues in the field of image dehazing: 1. **Insufficient dataset scale and diversity**: Existing dehazing datasets are relatively small in scale and resolution, and lack diversity in different haze intensities. This limits the generalization ability of current methods when dealing with complex real-world scenarios. 2. **Poor adaptability of models to different haze intensities**: Existing dehazing methods perform poorly when handling images with different haze intensities, especially when encountering previously unseen haze intensities. 3. **Lack of evaluation for downstream tasks**: Existing datasets rarely provide semantic annotations, making it difficult to evaluate the impact of dehazing methods on downstream tasks (such as object detection, semantic segmentation, and image description). To address these issues, the authors propose the following contributions: - **LMHaze Dataset**: This is a large-scale, high-resolution real-world dehazing dataset containing over 5,040 pairs of hazy and haze-free images, covering various indoor and outdoor scenes as well as different haze intensities. Additionally, the dataset provides multiple types of annotations, including labels for object detection, semantic segmentation, and image description. - **Benchmark Study**: Based on the LMHaze dataset, the authors conducted a comprehensive benchmark study, evaluating the performance of current state-of-the-art dehazing methods under different evaluation metrics (such as PSNR, SSIM, LPIPS, and metrics based on large modal models), and further assessed the performance of these methods in downstream vision tasks. - **MoE-Mamba Model**: The authors propose a new dehazing baseline method—the MoE-Mamba framework. This framework achieves robust dehazing performance for different haze intensities through three key components: - **Intensity-aware module based on large modal models**: Estimates haze intensity without the need for additional intensity labels. - **Mixture of Experts (MoE) module**: Dynamically adjusts network parameters based on the estimated haze intensity to handle images with different haze intensities. - **State Space Model (SSM) module**: Explores valuable non-local information while maintaining linear computational complexity. Through these contributions, the authors aim to improve the performance of image dehazing methods in real-world scenarios and promote further research in this field.

LMHaze: Intensity-aware Image Dehazing with a Large-scale Multi-intensity Real Haze Dataset

Nighttime Image Dehazing by Render.

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

Photo realistic synthetic dataset and multi-scale attention dehazing network

Dense Haze: A benchmark for image dehazing with dense-haze and haze-free images

SimHaze: game engine simulated data for real-world dehazing

A SINGLE IMAGE DEHAZING DATASET WITH LOW-LIGHT REAL-WORLD INDOOR IMAGES, DEPTH MAPS AND INFRARED IMAGES

From Synthetic to Real: Image Dehazing Collaborating with Unlabeled Real Data

Single image mixed dehazing method based on numerical iterative model and DehazeNet

I-HAZE: a dehazing benchmark with real hazy and haze-free indoor images

NH-HAZE: An Image Dehazing Benchmark with Non-Homogeneous Hazy and Haze-Free Images

A New Multi-Level Hazy Image and Video Dataset for Benchmark of Dehazing Methods

HazeSpace2M: A Dataset for Haze Aware Single Image Dehazing

Haze-Aware Attention Network for Single-Image Dehazing

Discrete Haze Level Dehazing Network

Robust Unpaired Image Dehazing Via Density and Depth Decomposition

Integrating Haze Density Features for Fast Nighttime Image Dehazing

Perceiving and Modeling Density is All You Need for Image Dehazing

Depth-agnostic Single Image Dehazing

4K-HAZE: A Dehazing Benchmark with 4K Resolution Hazy and Haze-Free Images

Multiscale Depth Fusion With Contextual Hybrid Enhancement Network for Image Dehazing