Uf-PromptGAN: A Nighttime Flare Removal Network Using Learnable Prompts

Flares are unwanted perturbances in a digital image surrounding a light source that hide details of the artifacts behind them. Many efficient data driven approaches using CNN and transformer based architectures have been recently proposed for the removal of reflective as well as scattering flares, but most of these networks work well only for daytime flares. For Night time flare removal, state of the art models use transformer based U-Net Architecture (Uformer), but they have not achieved truly accurate results. These models either fail to effectively remove flares or struggle to accurately reconstruct the artifacts hidden by them. In this paper, we introduce PromptGAN, a two stage model which first uses the Uformer to remove the flares followed by a learnable prompt-based AOT-GAN to inpaint the regions adversely affected by Uformer. We use learnable prompting to learn and encode background information images which helps the GAN to make more informed decisions while inpainting. Thus, the proposed approach ensures an effective removal of surrounding flares, resulting in a more realistic image reconstruction and accurate recreation of artifacts which were previously hidden by the flares.