MambaLE: a simple encoder-decoder architecture for low-light image enhancement with state space model

Low-light images is typically suffer from insufficient illumination, leading to blurred objects, loss of detail, reduced clarity, and muted colors. Enhancing the low light images is a vary crucial task in image processing by deep learning way. While recent methods have focused on improving enhancement quality by adjusting contrast, brightness, and compensating for both global and local features, existing dark enhancement frameworks often remain limited by biases inherent in local enhancement strategies. Selective state space models have recently demonstrated strong potential for capturing long-range dependencies with linear complexity, yet their application to low-light image deep learning enhancement has not been explored. In this work, we present MambaLE, a simple yet powerful benchmark model designed specifically for low-light image deep learning enhancement. Our approach introduces a decoder state space block as the core component, incorporates an efficient visual version state space block, and adds a convolutional fork branch to further boost performance. We formulate the enhancement task as a signal modulation problem, leveraging Haar wavelet transformation and attention mechanisms to augment the expressiveness of the standard state space model. We also investigate the integration of state space model blocks with convolutional operations. Experimental results on LOL benchmark low light enhancement datasets shows that MambaLE achieves outstanding quantitative and qualitative performance, highlighting its effectiveness and significance for low-light image deep learning way enhancement.