A Generative Adversarial Network with Effective Multiple-Scale Fusion of Features for Low-Light Image Inpainting

In normal lighting settings, most image-painting methods perform a great job of reconstructing image details and textures. However, in low-light environments, these subtleties are often ignored, leading to the loss of significant semantic information. To address these issues, we provide a single-stage generative adversarial network in this paper. Unlike previous methods, we fuse features generated from efficient multi-scale fusion at the coding layer in our encoder-decoder model. It does a good job of extracting image features to maintain finely detailed textures and improve the image’s overall brightness and clarity. Finally, we construct an adversarial network that is generative of two-branch discriminators to generate more realistic colors while preserving the image’s content. Extensive analysis and cross-dataset comparisons demonstrate the superior performance of our technology over existing methods.