卡通风格人脸图像生成研究

Study on Generation of Cartoon Style Face Images

通过风格迁移将卡通化的艺术表现形式施加到人脸图像上,能够获得个性化的卡通风格头像,现已成为计算机视觉的一个新研究方向。当前的风格迁移方法主要使用了生成对抗模型来学习图像的风格模式,虽然能够实现人脸图像的卡通化风格迁移,但容易引入一些瑕疵。在分析生成对抗模型工作原理基础之上,本文设计了使用注意力引导的生成对抗模型,通过注意力模型来引导生成对抗网络中的生成器,帮助其获得图像前景区域,从而降低风格迁移时对背景内容带来的影响。在将设计的模型应用到卡通人脸图像的生成时取得了非常好的性能,生成的头像很好地呈现了卡通艺术风格,同时也拥有非常好的视觉质量。

By applying the cartoon artistic presentation to face images,characteristic cartoon portraits can be acquired via image style translation.This has been a new research field in computer vision due to practical applications.Current approaches for image style translation generally employ the generative adversarial models to learn the style patterns first,and then they are used to generate target images.Though promising result has been achieved,they may produce visual artifacts.In this paper,the Generative Adversarial Networks(GAN)is deeply studied,and then a novel attention guided GAN model is proposed.With the help of non-local attention,the generator module in our model can identify the most discriminative foreground objects,which helps to minimize the impact on background.The proposed model is then applied to the generation of cartoon style faces,and pretty good effect is achieved.The generated images not only well present the cartoon art style,but also have favorable visual quality.