基于多特征融合的生成对抗网络图像修复算法  

Image inpainting algorithm of generative adversarial network based on multi feature fusion

作　　者：吴泓成 林国军 朱晏梅 王志舜 WU Hongcheng;LIN Guojun;ZHU Yanmei;WANG Zhishun(School of Automation and Information Engineering,Sichuan University of Science&Engineering,Yibin,Sichuan 644000,China;Artificial Intelligence Key Laboratory of Sichuan Province,Sichuan University of Science&Engineering,Yibin,Sichuan 644000,China)

机构地区：[1]四川轻化工大学自动化与信息工程学院,四川宜宾644000 [2]四川轻化工大学人工智能四川省重点实验室,四川宜宾644000

出　　处：《内江师范学院学报》2024年第12期39-45,共7页Journal of Neijiang Normal University

基　　金：四川省科技厅项目(2022YFSY0056);教育部产学合作协同育人项目(202102581011)。

摘　　要：针对现有图像修复算法存在结构不一致和纹理模糊等问题,提出一种基于多特征融合的生成对抗网络图像修复算法.该算法在传统的生成器中引入结合坐标注意力的多特征融合模块(CAMFM)获取更大感受野和多尺度特征.此外,生成器设计为双编码结构并引入注意力对图像进行特征提取,在生成器中引入VGG19网络提取特征用于计算感知损失和风格损失.在CelebA数据集上进行验证,计算修复结果的峰值信噪比(PSNR)为28.75dB,结构相似性(SSIM)为0.938,Fréchet Inception距离(FID)为5.99.该算法与五种基准算法比较,在三个指标上均最优,证明该算法具有好的修复性能.Aiming at problems in existing image inpainting algorithms,such as inconsistent structure and blurred texture,an image inpainting algorithm of generative adversarial network based on multi feature fusion was proposed.This algorithm introduces a multi feature fusion module(CAMFM)that combines coordinate attention mechanism in traditional generators to obtain larger receptive fields and multi-scale features.In addition,the generator is designed with a dual encoding structure and introduces attention for image feature extraction.The VGG19 network is introduced in the generator to extract features for calculating perceptual loss and style loss.Verified on the CelebA dataset,the peak signal-to-noise ratio(PSNR)of the repair results is 28.75dB,the structural similarity(SSIM)is 0.938,and the Fréchet Inception distance(FID)is 5.99.Compared with the four benchmark algorithms,the algorithm proposed in the article showed the best performance in all three indicators,proving that the algorithm proposed in the article has good repair performance.

关 键 词：多特征融合 图像修复 生成对抗网络 双编码 

分 类 号：TP751[自动化与计算机技术—检测技术与自动化装置]