Unsupervised Low-Light Enhancement Method Based on Attention Mechanism

Enhancing low-light images has been a research topic in the field of image processing. Significant advancements have been made recently using deep learning based approaches in this filed. However, obtaining the paired low/normal-light images in the same scene used for training is difficult, and the datasets lacks real-time, making its application value difficult to be verified. Meanwhile, low-light enhancement methods often ignore the local characteristics of images, which may result in local over or under exposure. To address the problems, the paper proposes a unsupervised generative adversarial network-based approach for image enhancement, incorporating attention in channel and spatial of images to improve the model generalization. Specially, our method does not require paired images as for datasets. At the generator component, we first apply channel attention into each feature map during the downsampling stage to focus on the channel information, and following the upsampling stage, the spatial attention feature map of the input image is multiplied with the generated results.we also perform a weighted summation of the input and the generated results to produce more realistic images. In discriminator component, we adopt a global-local discriminator that can enhance local regions while improving global lighting. We validate our method through extensive experiments on various benchmarks, and the results outperform the current cutting-edge methods.