Facial Appearance Modifications using SKPCA-Derived Features Extracted from Convolutional Autoencoder’s Latent Space
Krzysztof Adamiak, Paweł Kapusta, Krzysztof Ślot · 2020
The presented paper proposes a method that enables image object appearance editing by modifying its appearance-related high-level attributes. First, attribute-related features get extracted from a latent representation of an image generator and then, their contents gets modified, which results in producing the assumed appearance alterations. Convolutional Autoencoder (CAE) has been adopted as an image manipulation framework and face appearance, characterized by four attributes: age, smile intensity, facial hair intensity and gender, was chosen for modifications. To extract attribute-related features from CAE's latent representation, Supervised Kernel Principal Component Analysis (SKPCA) was used, as this transformation is able to disentangle complex, nonlinear image-to-attribute relationships. The method has been evaluated using large-scale face dataset CelebA. Qualitative results show that realistically-looking appearance modifications can be obtained. To quantify plausibility of introduced modifications, face recognition experiments on altered face images were performed, delivering on average 95% classification accuracy, for twenty-six category dataset.