Automatic Makeup and Remover Algorithm

Abstract

We proposed a facial makeup and remover algorithm base on Generative Adversarial Networks. The algorithm used Deep Convolutional Generative Adversarial Nets as the basic structure, and then extracted the image style information as well as applied style transformation through Cycle Generative Adversarial Networks. In pretreatment, the study applied affine algorithm in the training set to constrained the movement of facial key points. Thus, the algorithm ensured that the features of figures were not affected in the style transformation. Furthermore, we applied style conversion in designated areas, so the automatic facial makeup and remover algorithm could be used in images of any size. It should be highlighted that the experiment is essentially different from traditional makeup algorithms which adjust chroma, lightness, or use stickers to add up makeup effect. This experiment is a successful attempt in automatic facial style transformation because all effects are unpredictably generated by neural networks while the traditional ones are predictable. We improved the former image-portrait style conversion to the style conversion of image itself, where different people can acquire distinct makeup looks instead of portraits lacking fidelity. We also sets up a dataset of hundreds superstars to train and judge the effect of makeup algorithms.