基于边缘先验融合动态门控特征的人脸图像修复

为解决现有人脸图像修复算法因无法提取动态特征和缺乏边缘先验信息导致修复大区域不规则破损时纹理模糊和结构扭曲问题，提出了基于边缘先验融合动态门控特征的人脸图像修复算法。首先，设计动态门控卷积模块动态提取破损区域特征，关联已知区域和缺失区域的有效特征，提升纹理细腻度；然后，设计动态门控边缘增强网络和U型编码纹理修复网络，边缘增加网络旨在获取边缘轮廓信息，为U型编码纹理修复网络提供结构先验约束；U型编码纹理修复网络采用UNet++网络融合多层特征以保证人脸修复图像结构和纹理一致性；最后，通过消融实验证明UNet++网络的有效性和通用性，并剪枝U型网络以选取适宜的人脸图像模型表征层进行缺失区域纹理重建，在CelebA-HQ人脸数据集上进行实验评估。实验结果表明：相较于主流算法，所提方法在SSIM上平均提升3.87%，PSNR平均提升3.79 dB，FID平均下降16.54%，能有效修复大区域不规则缺失面积，生成纹理清晰、结构合理的图像。

Face image inpainting algorithm based on edge prior fusion dynamic gating features

To address the problem of texture blurring and structural distortion when repairing large irregular damaged areas in existing facial image repair algorithms, due to their inability to extract dynamic features and lack of edge prior information, this paper proposed a facial image repair algorithm based on edge prior fusion and dynamic gated feature. Firstly, it designed a dynamic gated convolution module to dynamically extract features from the damaged area, linking effective features of the known and missing areas to enhance the delicacy of the texture. Then, it designed a dynamic gated edge enhancement network and a U-shaped encoding texture repair network. The edge enhancement network aimed to capture edge contour information, providing structural prior constraints for the U-shaped encoding texture repair network. The U-shaped encoder texture repair network adopted the UNet++ network to fuse multi-layer features to ensure the consistency of structure and texture in the repaired facial image. Finally, the effectiveness and universality of the UNet++ network were demonstrated through ablation experiments, and the U-shaped network was pruned to select the appropriate facial image model representation layer for texture reconstruction in the missing area. Experiments were carried out on the CelebA-HQ facial dataset for evaluation. The experimental results show that compared with mainstream algorithms, the proposed method improves SSIM by an average of 3.87%, PSNR by an average of 3.79 dB, and reduces FID by an average of 16.54%, thereby more effectively repairing large areas of irregular missing regions, and generating images with clear texture and reasonable structure.