LERANet: Low-light Enhancement Network based on Retinex and Attention

The performance of vision based applications is often limited by low-light imaging environments. While various methods have been proposed to enhance image contrast, noise is inevitably amplified. In addition, most methods result in over-enhancement in bright regions. In this paper, we consider a convolutional neural network, named as LERANet, to decouple a dark image into reflectance and illumination, which can thus enhance contrast and reduce noise. An attention module is also integrated in the network to avoid over-enhancement. Experimental results demonstrate the effectiveness of the proposed LERANet on noise suppression and detail preservation. In addition, both subjective and objective comparisons with state-of-the-art algorithms indicate the superiority of the proposed method.