| 引用本文： |
-
杨辉,程汉荣,王富平,等.融合多结构信息和GAN的遮挡人脸修复算法[J].电讯技术，2025，（8）：1196 - 1203. [点击复制]
- YANG Hui,CHENG Hanrong,WANG Fuping,et al.An Occluded Face Inpainting Algorithm Integrating Multi-structure Information and GAN[J].，2025，（8）：1196 - 1203. [点击复制]
|
|
| 摘要: |
| 针对现有人脸图像修复算法的修复结果仍存在边缘模糊和结构扭曲等问题，提出了一种基于多结构信息融合和生成对抗网络(Generative Adversarial Network,GAN)的遮挡人脸修复算法。该算法构建了修复模块和信息融合再生成模块：修复模块由两个基于门卷积的U-Net网络组成，可实现相互约束和相互引导，生成更精细边缘和图像修复结果；信息融合再生成模块使用双向门控特征融合模块和上下文特征聚合模块将生成的信息融合再生，实现对缺失部分的精准补全。对修复模块生成的完整人脸图像进一步采用双重鉴别器来优化生成器的参数，提高算法的修复效果。在公开数据集Celeba-HQ和FFHQ上的实验结果表明，相比于对比算法，该算法在结构相似性上平均提升1.8%，峰值信噪比平均提升2.1 dB，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。 |
| 关键词: 人脸修复 信息融合 生成对抗网络 多先验属性 门卷积 |
| DOI：10.20079/j.issn.1001-893x.241118001 |
|
| 基金项目:陕西省科技厅重点研发计划（2024SF-YBXM-663） |
|
| An Occluded Face Inpainting Algorithm Integrating Multi-structure Information and GAN |
| YANG Hui,CHENG Hanrong,WANG Fuping,ZHEN Li,WANG Wenjing |
| (School of Communications and Information Engineering,Xi’an University of Posts and Telecommunications,Xi’an 710121,China) |
| Abstract: |
| For the problems that the face image inpainting results of existing algorithms often suffer from problems such as edge blurring and structural distortion,an occluded face inpainting algorithm based on multi-structure information fusion and generative adversarial network(GAN) is proposed.The inpainting module and information fusion and regeneration module are built,in which the inpainting module is composed of two gated convolution-based U-Net networks to achieve constraints and guidance for each other and produce good edge and image inpainting information.The generated information is fused and regenerated using bidirectional gated feature fusion modules and context feature aggregation modules to achieve precise completion of missing parts in the information fusion and regeneration module.For the complete face images generated by the information fusion and inpainting module,the edge discriminators and original image discriminators are adopted to optimize the parameters of the generator,further improving the inpainting effect of the algorithm.Experimental results on the public datasets Celeba-HQ and FFHQ indicate that,compared with competing algorithms,this method achieves an average improvement of 1.8% in structural similarity and an average boost of 2.1 dB in peak signal-to-noise ratio.It can effectively repair large,irregular missing areas and produce images with clear textures and well-structured details. |
| Key words: face inpainting information fusion generation adversarial network multiple prior attribute gate convolution |
《电讯技术》编辑部 