Pixel-level guided face editing with fully convolution networks

Abstract

Face editing has a variety of applications, especially with the increasing popularity of photography using mobile devices. In this work, we argue that the performance of face image editing can be further improved by using semantic segmentation which marks each pixel with a label that indicates its corresponding facial part. To this end, we propose a deep learning based method for automatic pixel-level labeling on face images. Our approach achieves state-of-the-art labeling accuracy on publicly available datasets, at a significantly higher speed than existing labeling methods. Then we show how the label information can be applied to various face image editing applications, such as face smoothing, face cloning and face blending. Extensive experimental results demonstrate the effectiveness of our method in editing face images with convincing visual quality.