MASTER Assignment
Restoration of damaged face statues using deep generative inpainting model
Type : Master M-CS
Period: Jan, 2019- Aug, 2020
Student : Abraham Theodorus, (Abraham, Student M-CS)
Date Final project: August 18, 2020
Supervisors:
Abstract:
Face statues restoration can be considered as an image inpainting task where a model aims to replace damaged regions with semantically correct pixels. In this thesis, Generative Adversarial Networks (GANs) with several objective functions are investigated in order to perform the task. As a comparison, a traditional technique, namely the recursive PCA algorithm, is also involved. The experiment results indicate superior performance of GANs in reconstructing the damaged face statues compared to the recursive PCA. Specifically, GANs are able to restore more complex face attributes well, i.e., eyes and mouths, while the recursive PCA still manages to restore noses which are considerably simpler. FID score is used to evaluate the inpainting results and it is suitable to be adopted as an early-stopping criterion for training GANs. However, among the inpainting results induced by GANs, only subtle differences are observed despite some slight FID score lead by certain GAN-based models. Therefore, manual observation is still deemed necessary.