Spatially Face Manipulation on Autoencoder Space

• 1、School of Artificial Intelligence, Beijing University of Post and Telecommunication, Beijing 100876

Abstract：While the quality of autoencoder image reconstruction and the disentanglement of autoencoder image representation have improved tremendously in recent years, the ability to manipulate the output image by controlling the latent space which represents images is still limited. Manipulation on the specific region of an image is also lack of study.This paper presents two novel face editing strategies that allow changing the semantic information of any arbitrary regions of images by manipulating the spatially disentangled representations of face images. One presents a new normalization, adaptive region normalization (AdaRN) to allow representation collaging, the other shows that the principal components computed by patch Principal Components Analysis (patch PCA) has meaningful information. The principal components allow to edit the specific region of image and control its semantic information. It was based on a well-trained autoencoder network called swapping autoencoder proposed recently.The two strategies can edit face images over an arbitrary region using weak supervision on a well-trained model. Experiments on FFHQ dataset show that any arbitrary regions such as mouth, eyes and eyebrows can be edited naturally using our strategies. Extensive results on the FFHQ dataset suggest that our strategy can not only edit face images flexibly but also require less effort for image labeling and model training tasks.

keywords： AutoEncoder, face manipulation, feature collage, latent space

点击查看论文中文信息

点击收起