Linear-ResNet GAN-based anime style transfer of face images

Converting directly real-world images into high-quality anime styles using generative adversarial networks is one of the research hotspots in computer vision. The current popular AnimeGAN and WhiteBox are problematic when distortion image features, loss details on lines textures concerned. To address these problems, we introduce AnimationGAN. preserve images, use linear bottlenecks residual network, what more, also employ hybrid attention mechanism to capture salient information images. In addition, adopt optimized normalizations improve accuracy learning rate model. experimental results show that compared with Whitebox, proposed AnimationGAN has smaller FID cartoon(61.73), better IS(6.79) faster network training speed(405 s per epoch). summary, generated animation significantly improves line texture feature retention much speed.