Image Inpainting Using Wasserstein Generative Adversarial Imputation Network

Image inpainting is one of the important tasks in computer vision which focuses on reconstruction missing regions an image. The aim this paper to introduce image model based Wasserstein Generative Adversarial Imputation Network. generator network uses building blocks convolutional layers with different dilation rates, together skip connections that help reproduce fine details output. This combination yields a universal imputation able handle various scenarios missingness sufficient quality. To show experimentally, simultaneously trained deal three given by pixels at random, smaller square regions, and placed center It turns out our achieves high-quality results all scenarios. Performance evaluated using peak signal-to-noise ratio structural similarity index two real-world benchmark datasets, CelebA faces Paris StreetView. are compared biharmonic some other state-of-the-art methods.