基于深度条件适应网络的标签转移算法  

LABEL TRANSFER ALGORITHM BASED ONDEPTH CONDITIONAL ADAPTIVE NETWORK

作　　者：刘雪锋 李京忠[2] Liu Xuefeng;Li Jingzhong(Xuchang Digital Learning Engineering Technology Research Center,Xuchang 461000,Henan,China;School of City and Environment,Xuchang University,Xuchang 461000,Henan,China)

机构地区：[1]许昌市数字化学习工程技术研究中心,河南许昌461000 [2]许昌学院城市与环境学院,河南许昌461000

出　　处：《计算机应用与软件》2023年第8期312-320,共9页Computer Applications and Software

基　　金：河南省青年人才托举工程和许昌学院科研项目(2020HYTP012)。

摘　　要：针对目前无监督域自适应方法对噪声和域偏移非常敏感,提出一种基于深度条件适应网络的标签转移算法。利用Wasserstein距离来度量区域分布差异,有效解决了当邻域差异较大时梯度消失问题,从而获得更好的域适应性能。提出一种条件适应策略,以减少域分布差异,解决边缘适应方法中经常忽略的类别不匹配和类别先验偏差。进一步引入一种标签相关传递算法预测伪目标标签,提升算法的准确性和实用性。对标准领域应用基准进行全面的实验,实验结果表明,该算法能够有效提升对噪声和域偏移的鲁棒性,进一步强化了算法的自适应性能。Aimed at the problem that the current unsupervised domain adaptive method is very sensitive to noise and domain offset,a label transfer algorithm based on depth condition adaptive network is proposed.Wasserstein distance was used to measure the regional distribution difference,which effectively solved the problem of gradient disappearing when the neighborhood difference was large,so as to obtain better domain adaptability.A conditional adaptation strategy was proposed to reduce the difference of domain distribution and solve the class mismatch and class prior bias which were often ignored in edge adaptation methods.Furthermore,a label correlation transfer algorithm was introduced to predict pseudo target tags,which improved the accuracy and practicability of the algorithm.A comprehensive experiment was carried out on the application benchmark in the standard field.The experimental results show that the proposed algorithm can effectively improve the robustness to noise and domain offset,and further enhance the adaptive performance of the algorithm.

关 键 词：无监督域 深度条件适应网络 标签相关传输 自适应 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术] TP18[自动化与计算机技术—计算机科学与技术]