Unsupervised Single Image Dehazing Using Dark Channel Prior Loss (1812.07051v2)

Abstract: Single image dehazing is a critical stage in many modern-day autonomous vision applications. Early prior-based methods often involved a time-consuming minimization of a hand-crafted energy function. Recent learning-based approaches utilize the representational power of deep neural networks (DNNs) to learn the underlying transformation between hazy and clear images. Due to inherent limitations in collecting matching clear and hazy images, these methods resort to training on synthetic data; constructed from indoor images and corresponding depth information. This may result in a possible domain shift when treating outdoor scenes. We propose a completely unsupervised method of training via minimization of the well-known, Dark Channel Prior (DCP) energy function. Instead of feeding the network with synthetic data, we solely use real-world outdoor images and tune the network's parameters by directly minimizing the DCP. Although our "Deep DCP" technique can be regarded as a fast approximator of DCP, it actually improves its results significantly. This suggests an additional regularization obtained via the network and learning process. Experiments show that our method performs on par with large-scale supervised methods.

• The paper introduces a novel unsupervised method for single image dehazing that trains a deep neural network directly on real-world hazy images using the Dark Channel Prior as a loss function.
• This unsupervised technique overcomes limitations of supervised methods by avoiding the domain shift associated with synthetic training data, leading to more robust performance in real outdoor scenes.
• Quantitative results show the proposed method achieves performance comparable to state-of-the-art supervised models and provides a 6.5 dB PSNR improvement over the classical Dark Channel Prior approach.

Unsupervised Single Image Dehazing Using Dark Channel Prior Loss

The task of single image dehazing is pivotal in autonomous vision systems, which rely heavily on clear visual data to perform accurately. The paper authored by Golts, Freedman, and Elad presents a novel method for single image dehazing which addresses some of the limitations faced by existing methodologies. The proposed approach leverages the Dark Channel Prior (DCP) and implements an unsupervised training paradigm using deep neural networks (DNNs), circumventing the pitfalls of synthetic training datasets commonly employed by other techniques.

Summary of the Method

A significant challenge in current image dehazing techniques lies in the acquisition of high-quality paired datasets of hazy and clear images. While synthetic datasets based on indoor depth data are often used, they bear the risk of domain shift when applied to outdoor scenes. This undermines the practical utility of supervised DNN-based methods when dealing with real-world outdoor imagery.

The authors capitalize on the DCP, traditionally employed in prior-based methods, to facilitate a loss function that guides an unsupervised learning process. The proposed "Deep DCP" model thus trains on real-world outdoor images, optimizing network parameters through direct minimization of the DCP energy function. This approach eschews synthetic training data entirely, relying solely on authentic outdoor hazy images for model training. The architecture of choice is a Context Aggregation Network (CAN) that utilizes dilated convolutions to efficiently predict transmission maps. These maps are integral in reconstructing dehazed images.

Quantitative Performance

The proposed approach is quantitatively evaluated on several benchmarks, notably RESIDE's SOTS-outdoor. The results demonstrate parity or superiority in performance when compared to sophisticated supervised models trained on synthetic data. A noteworthy achievement of this method is the 6.5 dB improvement in PSNR over classical DCP—this significant gain underscores the quality enhancement and regularization effects imparted by the DNN model.

Practical and Theoretical Implications

By proving that unsupervised training via energy function minimization can yield state-of-the-art results, this research opens avenues for broader applicability across various vision applications, particularly where genuine paired data is scarce or impractical to acquire. The reliance on actual hazy images addresses domain adaptation challenges, potentially leading to robust deployment scenarios in outdoor environments.

Future developments could explore integrating this unsupervised methodology with other known energy functions to enhance performance or extend applicability. Furthermore, investigating the synergy between minimal supervision signals and this unsupervised framework could potentially balance data acquisition costs with model accuracy.

Moreover, this work has implications for reducing computational overhead in training and inference stages—contributing to the practical deployment of dehazing models in resource-constrained environments such as embedded systems in autonomous vehicles.

Conclusion

This paper marks a defining step in unsupervised learning methods for image dehazing, providing a fully unsupervised solution that surpasses certain supervised counterparts. By employing a well-established prior like the DCP within a deep learning framework, it successfully avoids the shortcomings associated with synthetic training datasets, particularly the domain shift, thereby delivering real-world application efficacy. This advance highlights the potential of energy function-based loss formulations in unsupervised deep learning, paving the way for further research and development in this promising area of AI.