一种自我正则映射的弱光图像增强方法.

Most of the existing learning-based low-light image enhancement methods are training models with paired data. Although they have achieved good results, they lose their original advantage in the absence of paired training sets. Moreover, in practice, it is almost impossible to obtain two images with different brightness from the same perspective in the same scene, so the model cannot be completely trained with paired data. Therefore, in order to avoid using paired data and make the model domain more adaptive, based on the Generative Adversarial Network, an unsupervised self-regularized attention mapping method for low-light image enhancement is proposed, which is called SAMGAN. This method not only uses the illumination information of the original image as self-regular mapping and grayscale image to enhance it, but also uses the feature self-feature preserving loss to retain the features and content of the original image. Not only can it be trained in the absence of low/ normal-light image pairs, but it can also be well extended to various real-world test images. A large number of experiments have proved that the proposed method is superior to many current methods in terms of visual quality and subjective user study. [ABSTRACT FROM AUTHOR]