首页|基于注意力机制和边缘生成的图像修复研究

基于注意力机制和边缘生成的图像修复研究

扫码查看
原文链接
  • 万方数据
  • 维普
针对图像中大面积缺失区域修复产生纹理结构不清晰和上下文语义内容不协调的问题,提出基于注意力机制和边缘生成的图像修复算法.以生成对抗网络作为整体架构,添加边缘生成网络和ECA注意力机制模块,旨在强调破损区域内部边缘的修复,重点突出感兴趣区域与图像内容整体的一致性;采用感知损失函数进一步约束图像内容的生成,使纹理细节修复更加清晰.在CelebA和Paris Street View数据集上进行测试,实验结果表明,算法在峰值信噪比和结构相似度两项指标上均有所提升,对图像修复实践具有一定的参考价值.
Study on image inpainting based on attention mechanism and edge generation
To solve the problem of unclear texture structure and incongruity of contextual semantic content in the image when repairing large missing areas,an image repair algorithm is proposed based on attention mecha-nism and edge generation.The algorithm takes generative abduction network as the overall architecture,and adds edge generation module and ECA attention mechanism.The purpose of the repair network is to emphasize the semantic information of the inner edge of the damaged area and the area of interest to ensure the overall consistency of the repair content.At the same time,the perception loss function is modified to further constrain the generation of image content,and the inpainting details are more realistic.This method applies experiments in the CelebA and Paris Street View data sets.The experimental results show that both peak signal to noise ra-tio and structural similarity index measurement have been improved,which offers reference for the practice of image restoration.

image inpaintingedge generationattention mechanismperceptual loss

岳贺飞、刘华明、王秀友、毕学慧

展开 >

阜阳师范大学 计算机与信息工程学院,安徽 阜阳 236000

图像修复 边缘生成 注意力机制 感知损失

安徽省高校自然科学研究重大项目阜阳师范大学高层次人才科研启动项目阜阳师范大学校级项目阜阳师范大学校级项目阜阳师范大学校级项目阜阳市校合作项目阜阳师范大学阜阳市产业链研究创新团队

KJ2020ZD462020KYQD0032RCXM2020012020FSKJ122021FSKJ01ZDSXHZ202103CYLTD202213

2024

宁德师范学院学报(自然科学版)
宁德师范学院

宁德师范学院学报(自然科学版)

影响因子：0.303
ISSN：2095-2481
年,卷(期)：2024.36(3)