基于双判别器生成对抗网络的遮挡人脸图像修复算法

Research on Inpainting of Occlusion Face Image Based on Double Discriminator Generative Dversarial Network

对于当前遮挡人脸图像修复中,大多存在修复后人脸图像不连续、纹理模糊及网络训练过中存在模型崩溃等问题,针对这些问题提出了一种基于双判别器生成对抗网络的图像修复方法。该方法在全局判别器的基础上引入局部判别网络,以保证局部修复结果与周围区域的一致性;将encoder-decoder结构的卷积神经网络作为生成器,并在层间加入跳跃连接,从而提高模型对结构信息的预测能力;在判别器中引入Wasserstein距离,并添加梯度惩罚来训练两个判别模型,最终利用泊松图像编辑得到更加真实自然的修复结果。在CelebA人脸数据集上进行验证,实验结果表明该方法相较于所对比的文献模型具有更好的修复效果。

For the current occlusion face image repair,there are most problems such as repaired face image discontinuity,texture blur and model crash in network training,an image repair method based on dual discriminator generation and adversarial network in view of these problems is proposed.This method introduces local discrimination on the basis of global discriminator to ensure the consistency of local repair results and surrounding areas,takes the convolutional neural network of encoder-decoder structure as the generator to improve the prediction ability of structural information between layers,introduces Wasserstein distance in the discriminator and adds gradient penalty to train two discriminant models,and finally obtains more real and natural repair results with Poisson image editing.Validated on the CelebA face dataset,the experimental results show that this method has a better repair effect compared to the comparative literature model.