基于GAN的低照度图像增强算法研究

Research on low-light image enhancement algorithm based on GAN

针对低照度图像增强问题,提出一种基于生成式对抗网络(generative adversarial networks,GAN)的循环式图像增强网络.引入无监督学习方式,通过降低循环一致性损失和对抗性损失,估计低照度图像的原始光照图;利用建立的图像增强模型公式,对光照不足环境下采集到的图像进行亮度等方面的增强.在人工合成低照度图像数据集和真实自然低照度图像数据集上,均进行了质化和量化评价.实验表明,与现有的一些图像增强方法相比,该方法具有更好的图像增强效果,能够由低照度图像复原出生动、清晰、直观、自然的高质量图像.

To improve the performance of low-light image enhancement,a cycle-image enhancement network based on generative adversarial networks GAN is presented in this paper,as well as the concept of unsupervised learning is introduced by paring cycle-consistency loss and adversarial loss to obtain better estimates of the illumination maps of the original low-light images.Based on the estimated illumination maps,images collected in the environment of insufficient illumination can be enhanced by using the established image enhancement model formula.This paper conducts comprehensive qualitative and quantitative assessments through experiments using both synthetic and real low-light image datasets.The experimental results show that,the image enhancement method proposed in this paper can outperform much other state-of-the-art image enhancement methods and reproduce vivid,clear,visual and natural high-quality images from low-light images.