DEEP FACE - On the Reconstruction of Face Images from Deep Face Templates

Abstract

The paper on “Reconstruction of Face Images from Deep Face Templates" presents a novel approach for face image reconstruction using deep learning techniques. The proposed method utilizes a pre-trained deep face template, which is a convolutional neural network (CNN) trained on a large-scale face dataset, as a prior to guide the reconstruction process. Specifically, the method solves an optimization problem that balances the fidelity to the input image and the similarity to the deep face template. Its then evaluated with the method on two face image datasets, and demonstrate that their method outperforms several state-of-the-art methods in terms of reconstruction quality, especially for images with large occlusions or low resolutions. Moreover, they show that the deep face template can capture high-level face attributes, such as pose, identity, and expression, which can be used for various face-related tasks, such as face recognition, attribute manipulation, and generation. Overall, the paper presents a promising direction for face image reconstruction using deep learning techniques, and highlights the potential of deep face templates for capturing and utilizing high-level face attributes.