基于条件生成对抗网络的人脸补全算法

Face completion algorithm based on condition generation adversarial network

针对人脸遮挡区域重建问题,提出一种基于条件生成对抗式网络(CGAN)的人脸补全算法。首先遮挡人脸先通过卷积神经网络(CNN)进行五官等脸部特征提取,并作为一种约束信息输入生成器和判别器中,其中生成器将遮挡区域进行重构,重构人脸再分别输入局部信息判别器和全局信息判别器中,结合损失函数,最终生成完整人脸。在Celeb A数据集上,将重构后人脸与原图进行相似度比较,结果表明:该算法能够生成更加贴近原图的人脸。

Identifying occluded faces has become a research hotpot for face image processing in recent years.Aiming at the problem of reconstruction of face occlusion area,a face completion algorithm based on conditional generative adversarial nets(CGAN)is proposed.Firstly,the masking of human face is performed firstly by facial features extraction of facial features such as facial features through convolutional neural network(CNN),and it is used as a constraint information input generator and discriminator,in which the generator reconstructs occluded regions,reconstructs human faces and then inputs local regions respectively.The information discriminator and the global information discriminator combine the loss function to generate a complete human face.On the Celeb A dataset,comparing reconstructed and original faces,results shows that the algorithm can generate faces closer to the original image.