FaceMaskInpainting

Abstract

Recent advances in computational power and efficiency have allowed for the training of increasingly larger and more accurate neural networks. Despite these developments, model performance is still greatly restricted by network size, and thus computational resources. Given the increasing demand for complex models and the environmental damage associated with training these models, there is a great need for network optimization. In this paper, we explore the problem of object recognition and inpainting by attempting to remove face masks from images. To solve this problem, which we call mask inpainting, we utilize the “Face Mask Lite” dataset and examine various architectures and deep learning paradigms, including convolutional and generative adversarial networks. With the former, we develop a promising data-efficient encoder-decoder model that successfully removes masks and inpaints viable facial features.

See full paper Here

See full code Here