基于改进循环对抗网络的异质人脸合成配准  

Heterogeneous face synthesis matching based on improved cycle adversarial network

作　　者：汪鸿浩 张俊然[1] WANG Hong-hao;ZHANG Jun-ran(College of Electrical Engineering,Sichuan University,Chengdu 610065,China)

机构地区：[1]四川大学电气工程学院,四川成都610065

出　　处：《计算机工程与设计》2021年第6期1735-1741,共7页Computer Engineering and Design

基　　金：四川大学-泸州市校地合作基金项目(省市级科研项目)(2017CDLZ-G27)。

摘　　要：针对异质人脸合成问题,根据生成对抗网络在图像合成问题中的优势,将改进的循环生成对抗网络(CycleGAN)应用于异质人脸的合成研究当中。该方法可以降低不同模态人脸图像之间因为特征结构不同导致合成困难的问题。收集照片和素描两种模态的人脸数据,在两种模态之间使用循环对抗网络进行图像生成,实现两种模态之间的转换。利用最小二乘损失函数和Smooth损失函数对原网络目标函数进行改进,可以提高训练过程的稳定性,改善图片的合成质量,分析权重系数的作用。实验结果表明,合成质量和精度较原算法均有所提高。Aiming at the problem of heterogeneous face synthesis,according to the advantages of generative adversarial networks in image synthesis problems,the improved cycle generation adversarial network(CycleGAN)was applied to the synthesis of he-terogeneous faces.The difficult synthesis situation due to different feature structures between face images of different modalities was improved.Face data of two modes of photos and sketches were collected,and a cyclic adversarial network was used to gene-rate images between the two modes,so as to realize the conversion between the two modes.Using the least squares loss function and Smooth loss function to improve the original network objective function can improve the stability of the training process,improve the quality of picture synthesis,and analyze the role of weight coefficient.Experimental results show that the synthesis quality and accuracy are improved compared with the original algorithm.

关 键 词：深度学习 循环对抗网络 损失函数 图像合成 异质人脸配准 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]