Lightweight deep learning for image inpainting

Abstract

This project aims to investigate how the performance of a lightweight image inpainting can be improved while dropping the use of the discriminator and adversarial loss that is very common in most inpainting models. In our project, we made use of a generator model that was adapted from a GAN and has already been proven that it can accomplish image inpainting tasks very successfully. We also implemented different loss functions from different research and even came up with our own loss functions namely convoluted losses. Experiments were carried out to determine how these loss functions interact with one another in hopes of improving the performance on image inpainting. Finally, we also investigated whether a single model that is trained on a dataset with multiple category of images (faces and landscapes) can perform just as well as models that are trained on only one category of dataset. Our research ultimately shows that there is a reason why GANs are still the preferred method in image inpainting task but our loss functions and hybrid datasets showed some promise in possibly driving new ways of approaching an inpainting task