Curricular Contrastive Regularization for Physics-aware Single Image Dehazing (2303.14218v2)

Abstract: Considering the ill-posed nature, contrastive regularization has been developed for single image dehazing, introducing the information from negative images as a lower bound. However, the contrastive samples are nonconsensual, as the negatives are usually represented distantly from the clear (i.e., positive) image, leaving the solution space still under-constricted. Moreover, the interpretability of deep dehazing models is underexplored towards the physics of the hazing process. In this paper, we propose a novel curricular contrastive regularization targeted at a consensual contrastive space as opposed to a non-consensual one. Our negatives, which provide better lower-bound constraints, can be assembled from 1) the hazy image, and 2) corresponding restorations by other existing methods. Further, due to the different similarities between the embeddings of the clear image and negatives, the learning difficulty of the multiple components is intrinsically imbalanced. To tackle this issue, we customize a curriculum learning strategy to reweight the importance of different negatives. In addition, to improve the interpretability in the feature space, we build a physics-aware dual-branch unit according to the atmospheric scattering model. With the unit, as well as curricular contrastive regularization, we establish our dehazing network, named C2PNet. Extensive experiments demonstrate that our C2PNet significantly outperforms state-of-the-art methods, with extreme PSNR boosts of 3.94dB and 1.50dB, respectively, on SOTS-indoor and SOTS-outdoor datasets.

• The paper introduces a curricular contrastive regularization method that leverages consensual contrastive samples and curriculum learning to constrain the representation space.
• It integrates a physics-aware dual-branch unit that decouples atmospheric light and transmission map features for improved interpretability.
• Experimental results show significant PSNR gains and enhanced structural detail preservation on both indoor and outdoor dehazing benchmarks.

Curricular Contrastive Regularization for Physics-aware Single Image Dehazing: A Summary

Image dehazing, as a core visual preprocessing task, presents considerable challenges due to the ill-posed nature of recovering visibility from hazy images. The paper "Curricular Contrastive Regularization for Physics-aware Single Image Dehazing" by Zheng et al. advances this field by introducing innovative methods to enhance both the interpretability and efficacy of dehazing models. The authors propose curricular contrastive regularization that leverages a consensual contrastive space and a physics-aware feature synthesis approach.

Problem Formulation and Contributions

The primary thrust of this research is to tackle two interconnected issues: the unconstrained solution space of contrastive dehazing and the lack of interpretability in feature representations. Traditional approaches often employ contrastive regularization using negatives that are loosely related to the positive clear image. This makes the representation space insufficiently constrained, leading to suboptimal dehazing performance. Concurrently, existing deep dehazing models frequently disregard the atmospheric scattering physics that fundamentally governs the hazing process, resulting in less interpretable model behaviors.

Zheng et al. introduce a novel strategy using consensual contrastive samples, where negatives are derived from the original hazy images and their restorations using existing dehazing techniques. Such samples are inherently more informative because they retain content and differ only by haze distributions. Additionally, a curriculum learning strategy is employed to adaptively balance the learning across easy, hard, and ultra-hard negatives, thereby optimizing the learning trajectory.

The authors further present a physics-aware dual-branch unit (PDU), which more accurately approximates the atmospheric conditions described by the scattering model at the feature level. By decoupling the factors of atmospheric light and transmission map in dual branches, the decomposed features better align with the physics model, thereby enhancing the interpretability of the dehazing process.

Experimental Insights

The proposed dehazing network, C$^2$PNet, surpasses current state-of-the-art methods by achieving substantial improvements in PSNR on both SOTS-indoor (3.94dB) and SOTS-outdoor (1.50dB) datasets. This significant performance boost is attributed to the effective regularization induced by the curricular contrastive approach and the physics-consistent feature extraction facilitated by PDU. Visual inspection of results confirms that C$^2$PNet not only excels in quantitative metrics but also preserves structural details and reduces color distortions more effectively than traditional and existing competitive methods.

Implications and Future Directions

This paper contributes a structured framework that leverages curriculum learning to enhance the robustness of contrastive learning in image dehazing. The curricular adjustment of negative sample difficulty represents a substantial advancement in how contrastive learning can be tailored to complex tasks beyond traditional applications. The physics-aware dual-branch unit introduces a promising avenue for integrating domain knowledge directly into feature extraction processes within convolutional networks.

Future work may explore the application of this curricular approach to other domains that suffer from similar ill-posed challenges, such as low-light or underwater image enhancement. Additionally, the exploration of self-supervised or unsupervised variants of this methodology could further extend its applicability. Moreover, real-world validation of C$^2$PNet on diverse conditions and improving the model towards overcoming practical limitations in dynamic atmospheric environments presents compelling avenues for future research.

In conclusion, this paper provides a methodically profound contribution to the field of image dehazing, delivering both theoretical insights and practical gains that set a precedent for subsequent work in this domain.