Pool-mamba: Pooling state space model for low-light image enhancement

Abstract

Mamba, with its advantages in long-distance modeling and computational efficiency, has been promptly applied in low-light image enhancement (LLIE). Nevertheless, Mamba faces two key issues when processing low-light images: (1) overexposure in scenes with uneven illumination due to the lack of multi-scale modeling; (2) insufficient local detail recovery, as its sequential operation weakens the perception of local lighting changes. To alleviate these problems, we propose a Pooling State Space Model (Pool-Mamba) by integrating the state space model with pooling techniques. First, we devise the Pyramid-pooling Mamba (PMamba) module, which leverages pyramid pooling to capture multi-scale information, effectively mitigating uneven exposure under varying lighting conditions. Next, the Axis-pooling Mamba (AMamba) module is proposed to model local correlations along specific spatial dimensions (height and width), generating more refined local representations and enhancing the model’s ability to adapt to local lighting variations and intricate details. Finally, we incorporate a Dual Gated Enhancement Module (DGEM) to strengthen the channel correlations between PMamba and AMamba, facilitating the integration of multi-scale and local features. Benchmark assessments demonstrate that Pool-Mamba surpasses current state-of-the-art (SOTA) methods, achieving superior quantitative evaluations and less distorted visual results.