摘要
现存的人脸图像修复方法,在图像大比例缺损或分辨率高的条件下合成的图像会出现图像纹理结构不协调和上下文语义信息不一致的情况,不能准确合成清晰的图像结构,比如眼睛和眉毛等。为解决此问题,提出一种基于归一化注意力模块(NAM)的三阶段网络(RLGNet)人脸图像修复方法,能加快模型收敛速度并降低计算成本。其中,粗修复网络对残损图像进行初步修复后,局部修复网络对残损图像局部进行细致的修复,基于归一化注意力模块的全局细化修复网络对残损图像整体进行修复,以增加图像的语义连贯性和纹理结构的协调性。该方法在CelebA-HQ数据集上进行实验,结果表明在掩码比例为20%~30%时PSNR达到30.35 dB,SSIM达到0.9269,FID为2.55,能够合成相邻像素过渡自然和纹理结构合理的人脸图像。
Existing face image inpainting methods,which synthesize face images under the conditions of large scale image deficiency or high resolution that will have inconsistent image texture structure and inconsistent contextual semantic information,and cannot accurately synthe⁃size clear image structures,such as eyes and eyebrows,etc.To solve this problem,This paper proposed a Rough-Local-Global Networks(RL⁃GNet)face image inpainting method based on a Normalization-based Attention Module(NAM),which can accelerate the convergence speed of the model and reduce the computational cost.Among them,the coarse inpainting network performs the initial repair of the residual image and then the local inpainting network performs the detailed repair of the residual image locally;the global refinement inpainting network based on the normalized attention mechanism performs the repair of the residual image as a whole to increase the semantic coherence and the coordi⁃nation of the texture structure of the image.The method proposed in this paper is tested on the CelebA-HQ dataset.The results show that the PSNR reaches 30.35 dB and the SSIM value reaches 0.9269 and FID is 2.55 at the mask ratio of 20%~30%,which can synthesize face images with a natural transition of adjacent pixels and a reasonable texture structure.
作者
周遵富
张乾
李伟
李筱玉
ZHOU Zunfu;ZHANG Qian;LI Wei;LI Xiaoyu(School of Data Science and Information Engineering,Guizhou Minzu University;Key Laboratory of Pattern Recognition and Intelligent Systems of Guizhou,Guiyang 550025,China)
出处
《软件导刊》
2023年第8期196-202,共7页
Software Guide
基金
贵州省研究生科研基金项目(黔教合YJSKYJJ[2021]121)
贵州民族大学校级科研项目(GZMUZK[2021]YB23)。
关键词
人脸图像修复
归一化注意力模块
三阶段修复网络
激活函数
face image inpainting
normalized attention module
three-stage inpainting network
activation function
