Low Illumination Image Enhancement Algorithm Based on Generative Adversarial Network

Traditional deep learning-based methods have achieved promising performance for low-illumination image enhancement. However, these methods usually need to be trained on the pair-wise datasets, which are difficult to collect. Moreover, most existing enhancement methods have the problems of imperfect enhancement effect and image noise in real low illumination image enhancement. To address this, a unsupervised generative adversarial network is designed for low-illumination image enhancement, which has no requirement of training on the pair-wise datasets. The proposed network consists of two subnetworks: attentional mechanism network and enhancement network. The attentional mechanism network is used to distinguish the low-light region from the bright region of the low-illumination image, and the residual enhancement network is used to enhance the image by combining with the global-local discriminator. By doing this, a low-illumination image can be well enhanced. Extensive experimental results show that the proposed method outperforms the baseline Enlighten-GAN and Cycle-GAN for low-light image enhancement.