Abstract
Deep learning models for image dehazing often struggle with significant domain gaps between training datasets and a lack of fine-grained detail preservation. This paper introduces CrossDehaze, a data-centric framework that focuses on cross-data vision alignment. By employing channel-wise gamma correction to bridge domain gaps and a "weak-to-strong" self-supervised augmentation strategy, the proposed method achieves state-of-the-art results. Notably, CrossDehaze is significantly more parameter-efficient than its predecessors, outperforming models over ten times its size.
Addressing the Domain Gap in Image Dehazing
While large-scale training is standard for image dehazing, simply combining multiple datasets often leads to sub-optimal performance. These datasets—ranging from ground-level natural images to aerial remote sensing imagery—exhibit diverse color distributions, brightness levels, and sensor characteristics. These domain gaps hinder a model's ability to generalize effectively.
Furthermore, many existing architectures fail to preserve fine-grained local details when processing complex haze. Instead of merely increasing model capacity (e.g., through massive Transformers), CrossDehaze focuses on aligning and augmenting data to learn richer, more robust representations.
Methodology: Alignment and Consistency
The CrossDehaze framework consists of two primary components designed to improve representation learning without requiring massive parameter counts.
1. Cross-Data External Alignment
To merge disparate datasets like the Natural Image Dataset (NID) and Remote Sensing Image Dataset (RSID), the authors utilize a channel-wise gamma correction method. By calculating mean RGB values across datasets, they derive gamma factors to align color and contrast distributions. This allows the model to treat multiple datasets as a cohesive training signal, bypassing the need for complex, multi-phase domain adaptation.
2. Internal Augmentation with Self-Supervised Learning (WS-SSL)
To capture local structural details, the model implements a "Weak-to-Strong" consistency loss. The dehazed output undergoes two forms of augmentation:
- Weak Augmentation: Random cropping of the image.
- Strong Augmentation: Gaussian blurring of the cropped segment.
The model is trained to minimize the L2 difference between these versions, forcing it to learn features invariant to blur and focus on structural consistency.
3. Efficient Architecture
The underlying network is a 5-layer U-Net structure featuring Dehaze Blocks with multi-head self-attention and a Feature Fusion module. This module uses global average pooling to dynamically weight feature maps, ensuring the most relevant information is prioritized during reconstruction.
Key Results and Performance
CrossDehaze consistently outperforms state-of-the-art (SOTA) models across multiple benchmarks:
- Superior Accuracy: On the NID benchmark, it achieved a PSNR of 31.14 dB, surpassing MixDehazeNet (29.74 dB). On RSID, it reached 25.44 dB, exceeding DEA-Net (24.79 dB).
- Extreme Efficiency: The model utilizes only 2.51 million parameters. In comparison, Trinity-Net uses 20.24M and UHD uses 34.54M. Despite being a fraction of the size, CrossDehaze delivers better visual clarity and higher Structural Similarity Index (SSIM) scores.
- Visual Quality: Qualitative results show sharper edges and more natural color reproduction, particularly in challenging "out-of-distribution" scenarios where haze is non-uniform.
Conclusion and Future Directions
CrossDehaze demonstrates that data alignment and consistency-based learning are powerful alternatives to simply scaling up model size. While the self-supervised augmentation strategy increases training time due to the additional processing of augmented crops, the resulting gains in generalization and efficiency are substantial.
Future work will focus on optimizing these augmentation strategies to reduce training overhead and exploring the application of cross-data alignment to other restoration tasks, such as image deraining and low-light enhancement.
Summary updated