基于对抗生成网络的人脸照片去网纹技术

Method for Eliminating Mesh of Face Photograph Based on Generative Adversarial Networks

【目的】研究修复带网纹遮挡的人脸照片,以有助于提高后续的人脸验证的准确性。【方法】基于深度学习的模型(针对修复网纹遮挡的人脸照片)在训练时都要求输入对应的网纹数据,但是在实际应用中要获得对应的网纹数据却是非常困难的。为了避免使用对应的网纹数据对人脸图像进行有效的修复,提出了分离对抗生成网络。该网络利用网纹遮挡的人脸照片与干净的人脸照片做像素差生成残差网纹,利用1个分离网络和3个判别网络进行图像修复。【结果】实验结果表明所提出的方法对于消除人脸图像中网纹遮挡有效。【结论】针对带网纹结构遮挡的人脸图像,在对应网纹数据缺失的情况下,通过分离对抗生成网络,依然可以取得很好的图像修复效果。

[Purposes]In the field of image restoration,it is very important to repair the face images with mesh occlusion for improving the accuracy of face verification.[Methods]At present,the model based on deep learning for repairing mesh face requires input of corresponding mesh data in training,but it is very difficult to obtain the corresponding mesh data in practical application.In order to avoid using the corresponding mesh data to effectively repair the face image,a disentangling generative adversarial net is proposed.The residual mesh used to replace the real mesh is generated by using the pixel difference between the occluded face photograph and the clean face photograph.The disentangling generative adversarial net consists of a disentangling network and three discrimination networks for image restoration.[Findings]The experimental results show that the proposed method is effective in eliminating the reticulation of face images.[Conclusions]For the occlusion of the mesh structure of the face image,agood repair effect can still be achieved without using the corresponding mesh data.