Face Attribute Transformation Based On ConStarGAN

Abstract

Many models are able to transform styles by input images, such as Variational autoencoder (VAEs) and Generative adversarial networks (GANs), which have recently been applied to image style and domain transfer. In this paper, we propose a method based on unified generative adversarial networks for multi-domain image-to-image translation (StarGAN) to solve face attribute transfer problem—ConStarGAN. Given a face image, our model can extract the region of interest and transform multiple attributes in this region while keeping other features unchanged. So as to minimize the impact factor on the generated image and make it look very realistic. In our model, we present new loss function. Then, the image is segmented to avoid the influence of background, illumination and other factors, and spectral normalization is used to improve the quality of generated images. Experimental compared with the stability of relevant GAN models. Results show that we proposed model has achieved good results in face attribute translation. Finally, the effect of the improved model is illustrated through the effect analysis experiment.