基于生成对抗网络的遮挡人脸图像修复的改进与实现

针对目前的遮挡人脸图像修复领域中遮挡部位与遮挡大小的限制或修复后人脸图像不够连贯等问题,提出一种改进的Wasserstein生成对抗网络(WGAN)方法来改善人脸图像的修复。将卷积神经网络作为生成器模型,并在对应层间加入跳跃连接来增强生成图像的准确性。在判别器中引入Wasserstein距离进行判别,并引入梯度惩罚来完善判别器。在CelebA人脸数据集与LFW人脸数据集上进行实验,结果表明该方法的修复效果良好。

IMPROVEMENT AND IMPLEMENTATION OF OCCLUSION FACE IMAGE INPAINTING BASED ON GENERATIVE ADVERSARIAL NETWORK

In view of some problems of current occluded face image inpainting,such as the limitation of occlusion and occlusion size or the obvious repair of facial image after repair and the inconsistency of face image after restoration,this paper proposes an improvement based on Wasserstein Generative adversarial network(WGAN)to improve face image inpainting.It used the convolutional neural network as a generator model,and added jump connections between the corresponding layers to enhance the accuracy of the generated image.The Wasserstein distance was introduced into the discriminator to discriminate,and the gradient penalty was introduced to perfect the discriminator.Experiments were carried out on the CelebA face dataset and the LFW face dataset.The results show that the proposed method has a good repair effect.