Night-to-Day Translation via Illumination Degradation Disentanglement

TL;DR

This paper proposes a degradation disentanglement module and a degradation-aware contrastive learning module to identify different degradation patterns in nighttime images and introduces the degradation-aware contrastive learning strategy to preserve semantic consistency across distinct degradation regions.

Abstract

Night-to-Day translation (Night2Day) aims to achieve day-like vision for nighttime scenes. However, processing night images with complex degradations remains a significant challenge under unpaired conditions. Previous methods that uniformly mitigate these degradations have proven inadequate in simultaneously restoring daytime domain information and preserving underlying semantics. In this paper, we propose \textbf{N2D3} (\textbf{N}ight-to-\textbf{D}ay via \textbf{D}egradation \textbf{D}isentanglement) to identify different degradation patterns in nighttime images. Specifically, our method comprises a degradation disentanglement module and a degradation-aware contrastive learning module. Firstly, we extract physical priors from a photometric model based on Kubelka-Munk theory. Then, guided by these physical priors, we design a disentanglement module to discriminate among different illumination degradation regions. Finally, we introduce the degradation-aware contrastive learning strategy to preserve semantic consistency across distinct degradation regions. Our method is evaluated on two public datasets, demonstrating a significant improvement in visual quality and considerable potential for benefiting downstream tasks.

Figures (6)

• Figure 1: Illustration of our motivation. (a) The disentanglement process leverages physical priors. (b) Vanilla structure regularization and the corresponding results. (c) The proposed disentangled regularization and the corresponding results.
• Figure 2: The first row displays nighttime images, while the second row shows the disentanglement results using 3-Cluster K-means. The third row presents the disentanglement results with 4-Cluster K-means, and the last row shows the results from our physics-informed disentanglement. The color progression from blue, light blue, green to yellow corresponds to the following regions: darkness, well-lit, light effects, and high-light, respectively.
• Figure 3: The overall architecture of the proposed N2D3 method. The training phase contains the physical prior informed degradation disentanglement module and degradation-aware contrastive learning module. They are utilized to optimize the ResNet-based generator which is the main part in the inference phase.
• Figure 4: The qualitative comparison results on the BDD100K dataset.
• Figure 5: The qualitative comparison results on the Alderley dataset.

Theorems & Definitions (1)

• Corollary 1: Proof in the supplementary material