ODCR: Orthogonal Decoupling Contrastive Regularization for Unpaired Image Dehazing

TL;DR

The paper tackles unpaired image dehazing (UID) by addressing the conflict between preserving haze-related information and removing haze. It introduces Orthogonal Decoupling Contrastive Regularization (ODCR), which orthogonally projects image features onto separate haze-related and haze-unrelated subspaces using Orthogonal MLPs trained on the Stiefel manifold, and employs a self-supervised Depth-wise Feature Classifier (DWFC) to assign channel-wise weights. A Weighted PatchNCE (WPNCE) loss then maximizes mutual information for corresponding components across hazy and clear domains while respecting the decoupled spaces. The approach yields strong generalization on RESIDE, NH-HAZE 2, and Fattal datasets, outperforming several unpaired baselines and demonstrating that orthogonal feature decoupling can resolve the UID learning challenges without cycle-consistency.

Abstract

Unpaired image dehazing (UID) holds significant research importance due to the challenges in acquiring haze/clear image pairs with identical backgrounds. This paper proposes a novel method for UID named Orthogonal Decoupling Contrastive Regularization (ODCR). Our method is grounded in the assumption that an image consists of both haze-related features, which influence the degree of haze, and haze-unrelated features, such as texture and semantic information. ODCR aims to ensure that the haze-related features of the dehazing result closely resemble those of the clear image, while the haze-unrelated features align with the input hazy image. To accomplish the motivation, Orthogonal MLPs optimized geometrically on the Stiefel manifold are proposed, which can project image features into an orthogonal space, thereby reducing the relevance between different features. Furthermore, a task-driven Depth-wise Feature Classifier (DWFC) is proposed, which assigns weights to the orthogonal features based on the contribution of each channel's feature in predicting whether the feature source is hazy or clear in a self-supervised fashion. Finally, a Weighted PatchNCE (WPNCE) loss is introduced to achieve the pulling of haze-related features in the output image toward those of clear images, while bringing haze-unrelated features close to those of the hazy input. Extensive experiments demonstrate the superior performance of our ODCR method on UID.

Figures (9)

• Figure 1: Illustration of how (a) the CUT-like methods and (b) ODCR work. CUT-like methods directly pull features of the query patch and positive patch close, leading to a contradiction in maximizing the mutual information between the two patches and dehazing. In ODCR, orthogonal decoupling are proposed to decouple image features to haze-related (describing haze level) and unrelated (describing non-haze information, such as semantic and texture) components. Then the mutual information between query and the positive patches in different feature spaces are maximized, thus avoiding the above contradiction.
• Figure 2: The pipeline of the proposed ODCR.
• Figure 3: The illustration of the geometric optimization on the Stiefel manifold.
• Figure 4: The structure of DWFC and the procedure for obtaining the heat-vector describing the haze relevance of features.
• Figure 5: Visual comparison of various dehazing methods on SOTS-indoor li2018benchmarking and SOTS-outdoor li2018benchmarking. Areas where our method works better are boxed out and zoomed in, or you can zoom in by yourself to get a better view.