欢迎访问《图学学报》 分享到：
图学学报

图学学报 ›› 2021, Vol. 42 ›› Issue (1): 44-51.DOI: 10.11996/JG.j.2095-302X.2021010044

• 图像处理与计算机视觉 • 上一篇    下一篇

关键人脸轮廓区域卡通风格化生成算法

  

  1. 1. 温州大学计算机与人工智能学院，浙江 温州 325035；  2. 温州大学大数据与信息技术研究院，浙江 温州 325035
  • 出版日期:2021-02-28 发布日期:2021-02-01
  • 基金资助:
    国家重点研发计划项目(2018YFB1004904)；温州市科技计划项目(G20180036，R20200025) 

Generative adversarial network-based local facial stylization generation algorithm 

  1. 1. College of Computer Science and Artificial Intelligence, Wenzhou University, Wenzhou Zhejiang 325035, China;  2. Institute of Big Data and Information Technology of Wenzhou University, Wenzhou Zhejiang 325035, China
  • Online:2021-02-28 Published:2021-02-01
  • Supported by:
    The National Key Research and Development Program of China (2018YFB1004904); Basic Science and Technology Project of Wenzhou (G20180036, R20200025) 

PDF (PC)

123

摘要： 针对人脸轮廓特征区域的局部化限定，结合关键特征点的提取和脸部邻近颜色区域的融合，并 引入注意力机制，提出了一种基于 CycleGAN 的关键人脸轮廓区域卡通风格化生成算法，以此作为初始样本构 建生成对抗网络(GAN)并获取自然融合的局部卡通风格化人脸图像。利用人脸轮廓及关键特征点进行提取，结 合颜色特征信息限定关键人脸风格化区域，并通过局部区域二值化生成关键区域人脸预处理的采样图像；为了 使生成图像能够自然匹配所提取特征区域，利用均值滤波操作对所提取区域的边缘轮廓进行平滑羽化处理，并 相应地扩展风格化生成图像的过渡区域；最后通过构建基于无监督学习的生成对抗网络，使用训练数据集进行 人脸图像局部轮廓特征区域的卡通风格化生成。算法对人脸轮廓区域的边缘及邻近区域颜色进行滤波处理，可 实现良好的边缘轮廓过渡融合，生成自然的人脸局部轮廓区域的卡通风格化图像。实验结果表明，该算法对于 人脸图像的生成具有很高的鲁棒性，能够应用于各种尺度人脸图像的风格化生成。

关键词: 人脸特征, 局部区域, 对抗生成网络, 风格化

Abstract: In view of the localized facial contour features, combining with the extraction of key feature points and the fusion of adjacent color regions of the face, we presented a CycleGAN-based local facial stylization generation algorithm, and constructed the deep learning network with the attention mechanism to generate the local facial cartoon stylization. The sample facial images were marked by using the local area binarization method to constrain the key features and points. In order to naturally match the generated image with the extracted features, the mean filtering operation was utilized to smooth and feather the edge contour of the extracted region. Finally, the generative adversarial networks (GAN) network was constructed, and the training data set was employed to generate cartoon stylization images in the local contour feature area of facial images. The experiment results show that the presented algorithm exhibits high robustness for generating facial stylization, and that it can be applied to the generation of stylized facial images of various scales. 

Key words:  , facial features, local area, generative adversarial networks, stylization 

中图分类号: