| 摘要: |
| 本研究针对现有图像修复方法不能有效地分离结构和纹理信息,修复结果往往会出现边界模糊、结构扭曲 等伪影问题,提出了基于潜在特征重构和注意力机制的人脸图像修复方法。 人脸图像修复方法分为两阶段,第一 阶段,通过结构重建器网络提取样式向量,按照 StyleGAN 所述的原理分为粗尺度特征、中尺度特征和精细特征三 组,插入到预先训练好的 StyleGAN 生成器中,产生初步的修复结果;第二阶段通过构建纹理生成网络并使用上下 文注意力机制,注意力分数由注意力计算模块计算,注意力转移模块根据较高级别特征图和注意力分数来填充较 低级别特征图中的对应缺失区域,以细化上一阶段初步的人脸修复结果。 在 CelebA-HQ 数据集上的训练并进行 测试,本文的方法在定量和定性分析两个方面均优于现有方法。 因此,基于潜在特征重构和注意力机制的人脸图 像修复方法能够有效地修复缺损人脸图像,大大减少了边界过度平滑和存在纹理伪影的问题。 |
| 关键词: 图像修复 结构重建 纹理生成 注意力 |
| DOI： |
| 分类号: |
| 基金项目: |
|
| Face Image Restoration Based on Latent Feature Reconstruction and Attention Mechanism |
|
ZHANG Tongtong, LIU Heng
|
|
School of Computer Science and Technology, Anhui University of Technology, Anhui Maanshan 243000, China
|
| Abstract: |
| In this study, a face image restoration method based on latent feature reconstruction and attention mechanism was proposed to address the problem that existing image restoration methods cannot effectively separate structure and texture information, and the restoration results often show artifacts such as blurred boundaries and distorted structures. The face image restoration method was divided into two stages. In the first stage, the style vectors were extracted through the structural reconstruction network and divided into three groups of coarse-scale features, medium-scale features, and fine features according to the principles described by StyleGAN, which were inserted into the pre-trained StyleGAN generator to produce initial restoration results. In the second stage, by building a texture generation network and using a contextual attention mechanism, the attention score was calculated by the attention calculation module, and the attention transfer module filled in the corresponding missing regions in the lower-level feature images based on the higher level feature images and the attention scores to refine the initial face restoration results from the previous stage. Trained and tested on the CelebA-HQ dataset, the method in this paper outperformed existing methods in both quantitative and qualitative analysis. Thus, the face image restoration method based on latent feature reconstruction and attention mechanism can effectively repair defective face images, greatly reducing the problems of excessive smooth boundaries and the presence of texture artifacts. |
| Key words: image restoration structural reconstruction texture generation attention |
