针对人脸图像修复的深度学习网络存在修复后的人脸图像面部语义信息不合理和面部轮廓不协调的问题，提出了一种基于人脸结构信息引导的人脸图像修复网络。首先，采用编码器-解码器网络技术构建人脸结构草图生成网络，并在结构草图生成网络的生成器中加入跳跃连接和引入带膨胀卷积的残差块以生成待修复区域的结构草图。其次，在构建人脸修复网络时，在修复网络生成器中引入注意力机制，让修复网络在修复过程中更多关注待修复区域，并以生成的人脸结构草图为引导从而实现人脸图像面部语义结构和纹理信息的生动修复。最后，在结构草图生成网络的损失函数中引入特征匹配损失进行模型训练，从而约束生成器生成与真实结构草图更相似的结果；在修复网络的损失函数中联合感知损失和风格损失进行模型训练，从而更好地重建待修复区域的人脸图像面部轮廓结构和颜色纹理，使修复后的图像更接近真实图像。对比实验结果表明，在人脸图像数据集中，本文所设计的网络模型的修复性能有较高的提升。

Aiming at the problems of unreasonable facial semantic information and inconsistency of facial contours in the restored face image in the deep learning network for face image inpainting, a face image inpainting network guided by face structure information is proposed. Firstly, the encoder-decoder network technology is used to build a face structure sketch generation network, and skip connections and residual blocks with dilated convolution are added to the generator of the structure sketch generation network to generate the structure sketch of the region to be repaired. Secondly, when a face inpainting network is builted, an attention mechanism is introduced into the inpainting network generator, so that the inpainting network pays more attention to the area to be repaired during the inpainting process, and uses the generated face structure sketch as a guide to realize the face image vivid inpainting of facial semantic structure and texture information. Finally, the feature matching loss is introduced into the loss function of the structure sketch generation network for the model training, so as to constrain the generator to generate results more similar to the real structure sketch. In the loss function of the repair network, the perceptual loss and style loss are combined for the model training, therefore, the facial contour structure and color texture of the face image in the area to be repaired can be better reconstructed, so that the repaired image is closer to the real image. The comparative experimental results show that in the face image dataset, the repair performance of the network model designed in this paper has a high improvement.