TL;DR
This paper tackles the challenge of domain shift in image dehazing datasets by proposing a novel "cross-data vision alignment" approach. By aligning different datasets externally and using self-supervised learning internally, they achieve significantly improved dehazing performance, producing clearer images closer to real haze-free scenes.
What Problem Does This Paper Address?
Deep learning models for image dehazing have seen great success, but they often struggle when trained on diverse datasets. Previous methods typically combine multiple datasets to boost performance, but they tend to ignore the "domain gap" – the inherent differences in image characteristics (like color distribution and haze intensity) between these datasets. This oversight can lead to models that perform poorly, even when trained on large amounts of data. The authors aim to bridge this gap by developing a method that effectively aligns data from different sources, making models more robust and generalizable.
Methods
The core of the proposed method lies in a two-pronged approach:
- External Data Augmentation (Cross-Data Vision Alignment): This step addresses the domain gap between different datasets. The authors propose an "external augmentor" that aligns auxiliary datasets (like RSID) with the target dataset (like NID) in terms of color channels and image characteristics. This is achieved through gamma correction, which adjusts the brightness and contrast of images to match the target domain's distribution. This process ensures that the model learns from a more consistent and diverse set of data.
- Internal Data Augmentation (Self-Supervised Learning): To further enhance the model's ability to capture fine-grained details, an "internal augmentor" is employed. This uses a "weak-to-strong" self-supervised learning strategy. The model generates two versions of an image: one with weak augmentation (e.g., random cropping) and another with strong augmentation (e.g., Gaussian blur). By learning to minimize the difference between these two augmented versions, the model focuses on preserving local details and improving contrast, leading to more robust feature representations.
The overall framework combines these two augmentation strategies, feeding the enhanced data into a U-Net-based dehazing network.
Datasets
The paper uses the following datasets:
- NID (Natural Image Dataset): Used as the primary target dataset for training and evaluation.
- RSID (Remote Sensing Image Dataset): Used as an auxiliary dataset for external augmentation.
The authors select specific subsets of these datasets for training and testing (e.g., NID100 for testing).
Key Insights and Ideas
- Domain Gap is Crucial: The paper strongly emphasizes that simply concatenating diverse datasets without addressing their inherent differences (domain gap) leads to suboptimal performance in image dehazing.
- External Alignment for Robustness: Aligning datasets externally, particularly through color space adjustments like gamma correction, significantly improves the model's ability to generalize across different data sources.
- Internal Augmentation for Detail Preservation: Self-supervised learning, specifically the weak-to-strong approach, is vital for capturing and preserving fine-grained details that are often lost in traditional dehazing methods.
- Synergy of External and Internal: The combination of both external and internal augmentation strategies yields the best results, demonstrating a powerful synergy for robust and high-quality image dehazing.
Key Results
The proposed method achieves state-of-the-art results on both NID and RSID datasets.
- Quantitative Improvements: On the NID dataset, the method achieves a PSNR of 31.14 dB and an SSIM of 0.9896, outperforming existing methods like Trinity-Net and DEA-Net. On the RSID dataset, it also shows significant improvements, with a PSNR of 25.44 and SSIM of 0.9399.
- Qualitative Improvements: The dehazed images produced by the proposed method are visually closer to real haze-free images, with better texture details, clearer edges, and no color distortion compared to other methods. Figures 6, 7, 8, 9, 10, and 11 visually demonstrate these improvements.
- Efficiency: While the method has a higher runtime and computational complexity compared to some methods, it achieves superior metrics with fewer parameters, indicating good practical usability.
Limitations and Future Work
- Training Time: The authors acknowledge that the introduction of augmentation operations increases training time.
- Future Work: The paper suggests focusing on research aimed at improving training efficiency in the future.
Summary updated