摘要
为了克服几何信息不具有代表性和可区分性等缺点,提出一种基于几何感知双流网络的无监督域自适应模型。提出一种几何感知双流网络,从而实现跨域对齐和几何目标表示,在该网络中代表性信号通过对抗性适应网络获得;参考自适应几何统一标准的差异损失用于无监督几何对齐,目标域几何体受到约束。在无监督几何对齐中,通过几何一致性损失忽略目标数据的任意映射。实验结果表明该方法能够在跨域识别应用中取得较好的效果。
In order to overcome the shortcomings of non-representative and distinguish ability of geometric information,an unsupervised domain adaptive method based on geometric aware dual stream network is proposed.A geometry aware dual flow network was proposed to realize cross domain alignment and geometric target representation.In this network,representative signals were obtained through adversarial adaptive network.Referring to the difference loss of the adaptive geometric unified standard,it was used for unsupervised geometric alignment,and the geometry of the target domain was constrained.In unsupervised geometric alignment,any mapping of the target data was ignored through the geometric consistency loss.The experimental results show that the proposed method can achieve good results in cross domain recognition applications.
作者
韩彦净
马米米
张淑莉
Han Yanjing;Ma Mimi;Zhang Shuli(College of Technology,Zhengzhou Technology and Business University,Zhengzhou 450000,Henan,China;Henan University of Technology,Zhengzhou 450000,Henan,China)
出处
《计算机应用与软件》
北大核心
2023年第7期203-214,共12页
Computer Applications and Software
基金
河南省优秀青年科学基金项目(212300410036)。
关键词
无监督
域自适应
几何感知双流网络
几何对齐
Unsupervised
Domain adaptation
Geometry aware dual stream network
Geometric alignment
