GM-MoE: Low-Light Enhancement with Gated-Mechanism Mixture-of-Experts

TL;DR

GM-MoE tackles the generalization gap in low-light image enhancement by introducing a gated-Mechanism Mixture-of-Experts framework that dynamically weights three specialized sub-networks within a U-Net–like architecture. A gated weight generation network assigns per-image weights, enabling adaptive collaboration among color restoration, detail enhancement, and advanced feature augmentation, guided by shallow feature extraction (SFEB) for multi-scale information. The approach achieves state-of-the-art PSNR and SSIM across five LLIE benchmarks, with notable cross-domain gains on real and synthetic datasets, and robust qualitative improvements in color fidelity and detail preservation. This framework promises practical impact for autonomous driving, remote sensing, and surveillance by enhancing robustness across diverse lighting conditions and scenes.

Abstract

Low-light enhancement has wide applications in autonomous driving, 3D reconstruction, remote sensing, surveillance, and so on, which can significantly improve information utilization. However, most existing methods lack generalization and are limited to specific tasks such as image recovery. To address these issues, we propose Gated-Mechanism Mixture-of-Experts (GM-MoE), the first framework to introduce a mixture-of-experts network for low-light image enhancement. GM-MoE comprises a dynamic gated weight conditioning network and three sub-expert networks, each specializing in a distinct enhancement task. Combining a self-designed gated mechanism that dynamically adjusts the weights of the sub-expert networks for different data domains. Additionally, we integrate local and global feature fusion within sub-expert networks to enhance image quality by capturing multi-scale features. Experimental results demonstrate that the GM-MoE achieves superior generalization with respect to 25 compared approaches, reaching state-of-the-art performance on PSNR on 5 benchmarks and SSIM on 4 benchmarks, respectively.

Figures (12)

• Figure 1: Given a low-light image, our GM-MoE achieves better performance (for both object and whole scene) compared with LightenDiffusion jiang2024lightendiffusionunsupervisedlowlightimage.
• Figure 2: The comparison results among GM-MoE and the SOTA low-light image enhancement methods on the LOL-v1 , LOLv2-Synthetic and LSRW-Huawei benchmarks. GM-MoE outperforms all of compared approaches on both PSNR and SSIM metrics.
• Figure 3: Overview of the proposed GM-MOE. (a) The GM-MoE module comprises a gated weight generation network and three specialized sub-expert networks. (b) The overall network adopts a U-Net-like encoder-decoder architecture. Given an input image, it first undergoes processing through the Shallow Feature Extraction Block (SFEB). Then, the GM-MoE module facilitates multi-scale feature fusion via multiple downsampling and upsampling operations, ultimately generating the enhanced output image.
• Figure 4: Shallow Feature Extraction Block. The architecture of SFEB uses parallel convolutional and GAP for multi-scale feature capture.
• Figure 5: Qualitative comparison on LOLv1 (first row) and LOLv2-Synthetic(second row) . It can be seen that the proposed method significantly improves image clarity, and the colours are closer to reality.