基于多级滤波思想的人脸去手势遮挡恢复方法研究  

Research on Facial Gesture Occlusion Removal and Restoration Method Based on Multi-stage Filtering Thought

作　　者：欧静 文志诚[1] OU Jing;WEN Zhicheng(Hunan University of Technology,Zhuzhou 412000,China)

机构地区：[1]湖南工业大学,湖南株洲412000

出　　处：《现代信息科技》2023年第9期94-98,共5页Modern Information Technology

摘　　要：针对当下大面积人脸图像修复过程中广泛存在的色彩差异及生成内容丢失主要面部辨识特征等问题,文章提出一种端到端的三分支修复网络结构——TreeNet。单阶段的生成网络结构简单清晰且参数量较少,其中,逐像素滤波操作有益于减少色彩误差,促进生成图像的细节表达,记忆转化模块促使网络在结构记忆存储器中搜寻相关结构信息以重建缺失纹理信息。测试结果表明,TreeNet可以生成五官结构更加清晰自然、细节色彩更加逼近原图的人脸信息。To solve the problems of color difference and loss of main facial recognition features in the process of large-scale face inpainting,this paper proposes an end-to-end three branch repair network structure TreeNet.The single stage generation network has a simple and clear structure with fewer parameters.Among them,pixel by pixelfiltering operation is beneficial for reducing color errors and promoting the detailed expression of generated images.The memory conversion module enables the network to search for relevant structural information in the structural memory storage to reconstruct missing texture information.The test results show that TreeNet can generate facial information with clearer and more natural facial features and more detailed colors that approximate the original image.

关 键 词：深度学习 生成对抗网络 人脸修复 像素滤波 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]