关键词: 低光照图像, 无监督, 生成对抗网络, 注意力模块

Abstract: For the problems of poor brightness, contrast and details of low-light images, common enhancement algorithms rely on paired dataset for training, but this dataset is difficult to obtain. Therefore, we propose an unsupervised generative adversarial network PS-GAN algorithm based on unpaired dataset. Firstly, we construct SU-Net network to enhance the multi-scale feature extraction ability of the model; and we design the SP attention module to solve the problems of colour deviation and detail loss; then, we add the illumination smoothing loss function, so that the network generates a smoother and more natural image; finally, the simulation experiments on five public datasets show that the NIQE of PS-GAN is very low, and the NIQE of PS-GAN is very high, show that the NIQE index of PS-GAN reaches 3.7848, which is better than the existing mainstream methods.

Key words:  low-light images, unsupervised, generative adversarial networks, attention module