Facial feature transfer based on self-recognition style-encoding face editing network
Pengyuan Zhang, Yining Gao, Hang Zou, Yuhang Xiao, Pengjian Yang · 2021
This paper proposes a face attribute transfer system based on a deep convolutional network and a generative adversarial network, Self-recognition Style-encoding Face Editing Network(SSFEN). The network only needs a face source image to edit the facial features of the image. The whole network consists of two modules: self-recognition style-encoding network and multi-styles transfer network.The self-recognition style-encoding network aims to learn the style features of all faces in the dataset according to the input initial face image and features, and through further analysis and processing of the deep convolutional network, it outputs the facial feature coding of the original image. The multi-styles transfer network draws on the idea of generating a adversarial network, and only trains one generator to complete the style transfer in multiple fields, and adopts the generator to complete the editing of each field of a single face source image. In the joint training of the self-recognition style-encoding network and the multi-styles transfer network, we mainly apply the multi-label and multi-class discrimination loss, the adversarial attribute loss, the style domain recognition loss and the cyclic loss of the target domain. In addition, the style domain isolation loss is proposed to reduce the mutual influence between various target domains when the face is edited in a single target domain, which increases the accuracy of facial feature editing.