一种改进深度卷积生成对抗网络的人脸分割方法  

Face segmentation based on improved deep convolution GAN

作　　者：刘柏森[1] 邓琛[1] 张雾琳 LIU Baisen;DENG Chen;ZHANG Wulin(College of Electrical and Information Engineering,Heilongjiang Institute of Technology,Harbin 150050,China;College of Physics and Electronic Engineering,Mudanjiang Normal College,Mudanjiang 157000,China)

机构地区：[1]黑龙江工程学院电气与信息工程学院,黑龙江哈尔滨150050 [2]牡丹江师范学院物理与电子工程学院,黑龙江牡丹江157000

出　　处：《黑龙江工程学院学报》2019年第5期1-6,共6页Journal of Heilongjiang Institute of Technology

基　　金：黑龙江省高校青年学术骨干支持计划(1254G051)

摘　　要：在智能算法领域,人脸识别是一个重要的算法部分,而人脸分割又是人脸识别的一个重要组成部分。提出一种基于改进的深度卷积生成式对抗网络的人脸分割方法,将端到端的图像变换模式应用于生成器中,利用生成器对人脸图像进行分割。深度卷积生成式对抗网络将卷积层加入到生成器和判别器,使得生成器通过反卷积产生图像,而文中在反卷积之前再加入卷积层,组合形成全卷积的网络结构,将生成器的图像生成功能扩展成为语义分割功能。同时,生成器的输入原图和输出标签的通道组合作为判别器的判别对象,通过判别器来评判分割水平,进一步提高分割的标签与输入原图的关联性。经过多次实验,验证此方法能有效分割人脸主要区域。In the field of intelligent algorithms,face recognition is an important part of the algorithm,and face segmentation is an important part of face recognition.In this paper,a face segmentation method based on improved depth convolution generation antagonistic network is proposed.End-to-end image transformation mode is applied to the generator to segment the face image.Deep convolution adversarial network adds convolution layer to generators and discriminators,which makes generators generate images by deconvolution.In this paper,before deconvolution,convolution layer is added to form a full convolution network structure,which extends the image generation function of generators to semantic segmentation function.Then,the channel combination of the generator s input original image and output label is used as the discriminant object to judge the level of segmentation and further improve the segmentation,and the relevance between labels and input primitives.Experiments show that the proposed method can effectively segment the main face regions.

关 键 词：深度学习 语义分割 人脸分割 生成式对抗网络 深度卷积生成式对抗网络 

分 类 号：TN391.41[电子电信—物理电子学]