湖南大学学报（自然科学版）2024，Vol.51Issue(12)：87-97,11.DOI:10.16339/j.cnki.hdxbzkb.2024286
EMMA注意力与多尺度融合下的图像修复
Image Inpainting with EMMA Attention and Multi-scale Fusion
摘要
Abstract
To address the problem of accurately inferring the content of missing regions in an image when they are closely related to the surrounding textures and structures,we propose a single-stage image inpainting model.The model first compresses,reconstructs,and enhances features through convolutional layers and the FastStage module,while self-attention and multi-layer perceptron are incorporated to capture contextual relationships among features.Furthermore,in order to enhance the attention and importance perception on features,we propose EMMA in the models,which avoids the shaking and oscillation during updating the model parameters,thereby improving the performance of the generator and the quality of the generated results.Lastly,we introduce a discriminator to evaluate the consistency between the inpainted image and the original image.The end-to-end experimental results conducted on CelebA,Places2,and Paris StreetView datasets demonstrate that,compared with classical methods,the inpainting results of this model exhibit better visual semantics,and it is capable of finely inpainting details,textures,and local features of images.关键词
图像修复/注意力机制/膨胀卷积/深度学习Key words
image inpainting/attention mechanism/dilated convolution/deep learning分类
信息技术与安全科学引用本文复制引用
魏赟,王璐璐,邬开俊,单宏全,田彬..EMMA注意力与多尺度融合下的图像修复[J].湖南大学学报（自然科学版）,2024,51(12):87-97,11.基金项目
甘肃省自然科学基金项目(23JRRA913),Natural Science Foundation of Gansu Province(23JRRA913) （23JRRA913）
全国高等院校计算机基础教育研究会项目(2023-AFCEC-039),Association of Fundamental Computing Education in Chinese Universities(2023-AFCEC-039) （2023-AFCEC-039）
