Face attribute editing network based on style-content disentanglement and convolutional attention

Abstract

Face attribute editing is a research hotspot in the field of computer vision, which aims to modify a certain attribute of a face image to generate a new face image. The current methods based on Generative Adversarial Networks (GAN) have attribute entanglement problems and the implementation process is relatively complicated. To this end, this paper proposes a face attribute editing network based on style-content disentanglement and convolutional attention. Adding convolutional attention (CAT) module to the StyleGAN generator makes the network's control of content features no longer affected by the overall style of the image, and realizes the separation of spatial content and style from coarse to fine. In addition, the hierarchical CAT modules control different levels of attribute features, and changing the input of any layer of CAT can change the corresponding attribute features. The experimental results on the CelebA-HQ dataset show that the method in this paper can achieve disentangled editing of face attributes, and the scores of various indicators are better than the existing models.