Abstract
Presence of haze in images obscures underlying information, which is undesirable in applications requiring accurate environment information. To recover such an image, a dehazing algorithm should localize and recover affected regions while ensuring consistency between recovered and its neighboring regions. However owing to fixed receptive field of convolutional kernels and non uniform haze distribution, assuring consistency between regions is difficult. In this paper, we utilize an encoder-decoder based network architecture to perform the task of dehazing and integrate an spatially aware channel attention mechanism to enhance features of interest beyond the receptive field of traditional conventional kernels. To ensure performance consistency across diverse range of haze densities, we utilize greedy localized data augmentation mechanism. Synthetic datasets are typically used to ensure a large amount of paired training samples, however the methodology to generate such samples introduces a gap between them and real images while accounting for only uniform haze distribution and overlooking more realistic scenario of non-uniform haze distribution resulting in inferior dehazing performance when evaluated on real datasets. Despite this, the abundance of paired samples within synthetic datasets cannot be ignored. Thus to ensure performance consistency across diverse datasets, we train the proposed network within an adversarial prior-guided framework that relies on a generated image along with its low and high frequency components to determine if properties of dehazed images matches those of ground truth. We perform extensive experiments to validate the dehazing and domain invariance performance of proposed framework across diverse domains and report state-of-the-art (SoTA) results.
Introduction
Single image dehazing is a critical preprocessing step for many computer vision applications. However, most existing deep learning models struggle with two primary issues: the uniformity assumption and the domain gap. Standard synthetic datasets often apply haze uniformly across an image, failing to represent the non-homogeneous nature of real-world fog or smoke. Furthermore, models trained on these "perfect" simulations often fail to generalize to real-world photography due to mathematical differences in atmospheric interference.
Methodology
The proposed framework, DIDH (Domain Invariant Dehazing), utilizes an encoder-decoder architecture enhanced by three key innovations:
Spatially Aware Channel Attention (SACA)
To overcome the limitations of fixed receptive fields in standard CNNs, SACA modules are integrated into the skip connections. These modules use non-local operations to capture long-range dependencies, allowing the network to maintain consistency between distant regions during restoration.
Copy-Blend Augmentation
To address non-uniform haze, a "greedy localized data augmentation" technique is employed. By blending hazy patches of random shapes into clean images, the model learns to localize and recover specific affected regions rather than applying a global filter.
Frequency Prior-Guided Adversarial Training
The training process incorporates two discriminators focusing on Low Frequency (LF) and High Frequency (HF) components. This disentanglement ensures that the model preserves both global structural/color consistency and fine-grained edge details.
Experimental Results
The model was evaluated on a diverse set of benchmarks, including synthetic (RESIDE, HazeRD) and real-world (NTIRE-2018/19/20) datasets.
- Real-World Superiority: On the NTIRE-19 dataset, DIDH achieved a PSNR of 22.18 dB, significantly outperforming MSNet (19.07 dB).
- Domain Invariance: When tested on NTIRE-18 without specific tuning, it reached 24.14 dB, whereas other state-of-the-art models typically stayed below 19 dB.
- Ablation Studies: Results confirmed that both the SACA modules and frequency-guided adversarial loss are essential for achieving domain invariance.
Conclusion
By moving away from global filters and uniform haze assumptions, the DIDH framework provides a robust solution for real-world dehazing. The combination of localized augmentation and frequency-aware training allows the model to bridge the gap between synthetic training data and the complexities of real-world atmospheric conditions. Future work will explore reducing reliance on paired data and optimizing the architecture for real-time video applications.
Summary updated