Abstract: Low-light image enhancement technology is a key method for improving the quality of image captured in low-light environments,thereby ensuring the reliability of subsequent visual tasks. To enhance the generalization capability of low-light image enhancement technology in unknown scenarios and to break through the limitations of traditional methods that rely on paired datasets for training,this paper proposes a low-light image enhancement approach based on a single-path generative adversarial network (GAN) combined with a multi-attention mechanism. Firstly,a multi-attention-guided generator network is constructed with a multi-dimensional attention mechanism to extract the inverted illumination features of images to guide the enhancement of low-light images. By improving the self-attention mechanism,the long-range dependencies among image pixels are enhanced to improve the network's generalization ability. Secondly,a dual-discriminator architecture is crafted to discern the authenticity of images from both global and local perspectives. Finally,the network training is regularized by incorporating self-feature preservation loss and the adversarial loss of GAN. Experimental results on unpaired datasets demonstrate that this method not only overcomes the reliance on paired datasets but also exhibits robust adaptability to low-light image enhancement tasks under complex conditions. It achieves superior visual quality and the best performance in both objective and subjective evaluations on public datasets such as MEF,LIME,and DICM.