第二十二届中国虚拟现实大会 (ChinaVR 2022)

54 / 2022-09-29 03:34:58

CartoonFace-CycleGAN: Cartoon Face Generation Based on Deep Learning

In this paper, we propose a Generative Adversarial Network to transform human face into anime style cartoon image. Although there are many well-known cartoon stylization algorithms in the field of image style transfer, few researches on the generation of cartoon face, since anime style has high requirements for the distribution of facial features and local area details. In order to generate high-quality cartoon face, this paper proposes a new network structure: CartoonFace-CycleGAN (CF-CycleGAN). We use classical CycleGAN network to train global transformation with unpaired images (i.e. face photo and cartoon image). On this basis, we add landmark cropped local discriminator and regional attentive cycle loss to train the local region of facial features, use eye-promoting global adversarial loss to maintain eye consistency, and apply high-level semantic loss to maintain the semantic information between the original image and the generated one. Experimental results show that our network can produce high-quality cartoon face while maintaining the diversity of output.