| 引用本文: | 石计亮1,2 ,张 乾1,2.结构与纹理双生成的二阶段网络图像修复(J/M/D/N,J:杂志，M：书，D：论文，N：报纸).期刊名称,2025，42（1）：9-19 |
| CHEN X. Adap tive slidingmode contr ol for discrete2ti me multi2inputmulti2 out put systems[ J ]. Aut omatica, 2006, 42(6): 4272-435 |
|
| 摘要: |
| 目的 针对现有图像修复方法不能很好地实现结构和纹理信息之间的双向交互,在修复缺失面积较大或纹
理复杂的图像时存在纹理模糊、结构失真等问题。 方法 提出了一种基于双向坐标注意融合模块和傅里叶特征聚合
模块的二阶段网络图像修复方法。 首先,使用结构编-解码器和纹理编-解码器对受损图像进行结构重建和纹理合
成,产生初步的修复结果;然后,将粗修复结果输入到细化修复网络,利用双向坐标注意融合模块和傅里叶特征聚
合模块对图像内部纹理细节进行修复;为增强全局一致性,设计了双向坐标注意融合模块来实现结构和纹理信息
之间的双向交互,并设计了傅里叶特征聚合模块,用于捕获全局上下文信息,增强图像局部特征之间的相关性,以
获得精细的修复结果;此外,还利用双流判别器来估计结构和纹理的特征统计量,以区分原始图像和生成图像。
结果 在 CelebA-HQ 数据集上进行实验,与 4 种图像修复方法进行比较,定性结果表明方法生成的人脸图像更加清
晰自然;定量结果表明方法在峰值信噪比、结构相似性指数和弗雷歇距离上均优于对比算法;对模型中各模块的消
融实验结果也验证了所提出创新点的有效性。 结论 因此,所提出的方法能够有效地修复受损的人脸图像,特别是
在大面积遮挡下也能生成具有结构合理、纹理清晰的图像。 |
| 关键词: 图像修复 二阶段网络 生成对抗网络 深度学习 |
| DOI： |
| 分类号: |
| 基金项目: |
|
| Two-stage Network Image Inpainting with Dual Generation of Structure and Texture |
|
SHI Jiliang1 2 ZHANG Qian1 2
|
|
1. School of Data Science and Information Engineering Guizhou Minzu University Guiyang 550025 China
2. Key Laboratory of Pattern Recognition and Intelligent Systems of Guizhou Province Guizhou Minzu University Guiyang
550025 China
|
| Abstract: |
| Objective Existing image inpainting methods fail to achieve effective bidirectional interaction between structure
and texture information resulting in issues like texture blur and structural distortion when repairing images with large
missing areas or complex textures. Methods A two-stage network image inpainting method was proposed employing a
bidirectional coordinate attention fusion module and a Fourier feature aggregation module. Firstly the damaged image was
subjected to structure reconstruction and texture synthesis using structure encoder-decoder and texture encoder-decoder
generating preliminary inpainting results. Subsequently the coarse inpainting result was input to a refinement inpainting
network where the bidirectional coordinate attention fusion module and the Fourier feature aggregation module were utilized to repair the internal texture details of the image. To enhance global consistency the bidirectional coordinate
attention fusion module was designed to facilitate bidirectional interaction between structure and texture information.
Additionally the Fourier feature aggregation module was designed to capture global contextual information enhancing the
correlation between local image features to obtain fine inpainting results. Moreover dual-stream discriminators were
employed to estimate the feature statistics of structure and texture distinguishing between original and generated images.
Results In experiments conducted on the CelebA-HQ dataset compared with four image inpainting methods qualitative
results indicated that face images generated by this method were clearer and more natural the quantitative results showed
that this method outperformed the contrastive algorithms in peak signal-to-noise ratio structural similarity index and
Fréchet distance. Ablation experiments on various modules of the model also validated the effectiveness of the proposed
innovations. Conclusion Therefore the proposed method effectively restores damaged face images especially generating
images with reasonable structure and clear texture even under large occlusions. |
| Key words: image inpainting two-stage network generative adversarial network deep learning |
