基于生成对抗网络的动漫风格迁移

Anime Style Migration Based on Generative Adversarial Networks

针对目前生成动画图像质量较低、内容失真、视觉效果有待进一步提升等问题,本文在Generative Adversarial Networks for Photo Cartoonization(CartoonGAN)网络的基础上提出了一种动漫风格迁移的网络模型。首先通过引入预训练模型ResNet-101对内容损失函数进行训练,加快训练速度的同时保证真实图片内容的完整性;然后在生成器模块加入激励挤压模块(SE-Block),实现在特征提取中保留空间特征和通道特征,使得生成的图片更易被判别器区分,从而更好地训练判别器。最后,进行了定性比较和定量分析,结果表明本文所提算法能够有效提升训练速度、提高漫画图像生成质量和增强图像的抽象感,且IS、FID的得分分别为11.2和86。

Aiming at the current issues of low image quality,content distortion,and room for improvement in visual effects in generated animated images,a network for anime style migration has been proposed based on the generative adversarial networks for photo cartoonization(CartoonGAN)network.The training speed of content loss function is enhanced by incorporating with the pre-trained model ResNet-101,maintaining the integrity of real image content at the same time.Additionally,an excitation-squeeze block module is added into the generator module to retain the spatial and channel features during feature extraction,making generated images easier to distinguish by the discriminator,which is preferable to train the discriminator.Finally,qualitative comparisons and quantitative analyses show that the network proposed in the present study can effectively increase training speed,improve generated anime image quality and enhance the abstraction of generated cartoon images.Moreover,the IS and FID scores are 11.2 and 86,respectively.