基于生成对抗网络的人脸转正方法研究

Research on face transformation method based on generative adversarial network

为了增加偏角人脸的特征信息,提升人脸识别准确率,研究一种基于生成对抗网络(GAN)的人脸转正方法。首先将GAN网络的生成器框架由单通道变为双通道结构,方便对全局结构和局部纹理进行处理,然后对判别器加入注意力机制,最后引入合成损失函数使其能够高效地生成人脸正面图片。实验结果证明:将人脸转正图片输入识别网络后,识别准确率有所提升,随着角度变化识别率衰减程度变小,能够极大程度保留人脸重要的识别信息,对后续的人脸识别工作有较大的帮助。

In order to increase the feature information of off-angle faces and improve the accuracy of face recognition,a face trans⁃formation method based on Generative Adversarial Network(GAN)is studied.First,the generator frame of the GAN network is changed from a single-channel to a dual-channel structure,which is convenient for processing the global structure and local texture,then an at⁃tention mechanism is added to the discriminator,and finally a synthetic loss function is introduced to enable it to efficiently generate the front of the face picture.The experimental results show that:after the face is turned into a positive image and input into the recogni⁃tion network,the recognition accuracy rate is improved,and the degree of attenuation of the recognition rate becomes smaller with the change of angle,which can greatly retain the important recognition information of the face,and it can be used for subsequent face recog⁃nition.Recognition work is of great help.