Nighttime Hazy Image Enhancement via Progressively and Mutually Reinforcing Night-Haze Priors

TL;DR

This work tackles the challenge of nighttime hazy image enhancement by jointly leveraging haze priors and low-light priors. It introduces M3KE, a Multi-level Mixture of Mutual Knowledge Experts that progressively combines image-, patch-, and pixel-level experts operating across visual and frequency domains, guided by a Frequency-Aware Router and integrated Frequency Interaction Blocks. The method explicitly models the mutual knowledge exchange between dehazing and low-light enhancement, achieving state-of-the-art results on nighttime dehazing benchmarks and demonstrating generalization to daytime dehazing and low-light enhancement tasks. The approach enables robust restoration under diverse degradations by leveraging cross-domain priors, resulting in improved visibility, natural color balance, and preserved details in real-world scenes.

Abstract

Enhancing the visibility of nighttime hazy images is challenging due to the complex degradation distributions. Existing methods mainly address a single type of degradation (e.g., haze or low-light) at a time, ignoring the interplay of different degradation types and resulting in limited visibility improvement. We observe that the domain knowledge shared between low-light and haze priors can be reinforced mutually for better visibility. Based on this key insight, in this paper, we propose a novel framework that enhances visibility in nighttime hazy images by reinforcing the intrinsic consistency between haze and low-light priors mutually and progressively. In particular, our model utilizes image-, patch-, and pixel-level experts that operate across visual and frequency domains to recover global scene structure, regional patterns, and fine-grained details progressively. A frequency-aware router is further introduced to adaptively guide the contribution of each expert, ensuring robust image restoration. Extensive experiments demonstrate the superior performance of our model on nighttime dehazing benchmarks both quantitatively and qualitatively. Moreover, we showcase the generalizability of our model in daytime dehazing and low-light enhancement tasks.

Figures (14)

• Figure 1: Comparisons with recent state-of-the-art methods (DCMPNet DCMPNet for daytime dehazing, ReDDiT ReDDiT for low light enhancement, NightHaze NightHaze for nighttime dehazing) and our method on diverse nighttime hazy scenarios. Conventional methods that address either daytime dehazing or low-light enhancement in isolation can only recover limited aspects of nighttime hazy images. In contrast, our method leverages both priors jointly, achieving more comprehensive image restoration and surpassing the performance of existing state-of-the-art nighttime dehazing techniques.
• Figure 2: The overall pipeline of our M3KE framework. Given a degraded image under diverse hazy or low-light scenarios as input, M3KE progressively enhance the visibility through image-, patch-, and pixel-level experts respectively. A Frequency-aware Router is employed to guide the weight assignment for each expert. Finally, a clear image is generated as the output.
• Figure 3: Illustration of the Expert Block, which is designed to fuse and enhance features from different domains.
• Figure 4: Qualitative comparison of the proposed method with prior works for synthetic nighttime image dehazing.
• Figure 5: Qualitative comparison of the proposed method with prior works for real-world nighttime image dehazing.