基于生成对抗网络的面部修复

如何更好地对受损的面部图像实施相应的修复,根据此问题指出了一类基于生成对抗网络改良以后的面部修复算法.首先,在生成模型中把编码器和解码器的中间层的全连接换成逐信道全连接,在编码和解码阶段使用卷积操作代替池化操作,针对损失函数采用的激活函数进行改进,增加tanh函数,提高图像补全效果.然后,在保证功能上不受损并且输入、输出尺寸保持原状的条件下对判别器的模型进行了相应的改良,最后,对损失函数引进TV损失、重建损失这二者来实现对生成网络的优化处理,由此提升细节图像方面的修复实力.通过实验表明,使用该方法修复后的面部图像,比先前的方法更清晰更连贯.

Facial Repair Based on Generative Adversarial Networks

How to better repair damaged facial images? Aiming at this problem,an improved facial repair algorithm based on generative adversarial network is proposed.First,in the generative model,the full connection of the middle layer of the encoder and decoder is replaced by the full connection of each channel.In the encoding and decoding stage,the convolution operation is used instead of the pooling operation.The activation function used for the loss function is improved and increased.The tanh function improves the image completion effect.Then,under the condition that the function is not damaged and the import and export dimensions remain the same,the discriminator model is improved accordingly.Finally,the TV loss and reconstruction loss are introduced into the loss function to jointly optimize the generation network and enhance the detailed image.Repair ability experiments show that the facial images restored using this method are clearer and more consistent than the previous methods.