Advance of image processing APPs gives rise to an overwhelming trend of artiﬁcial beautiﬁcation. Our project aims to reverse the portrait beautiﬁcation despite lack of information of the exact process applied. To precisely capture the subtle features of facial images, we apply Principle Component Analysis to the residual images of the original and beautiﬁed images, and then train several independent convolutional neural networks to regress each normalized components, and finally restore the original faces.
We used the web crawler and portraitbeautifying APPs to build our own dataset.
We re-constructed and improved the Component Regression Network proposed in this paper. Generally, we regressed the residual image between the beautified one and the original one in the scaled-component space.