基于可调节判别器的领域适应  

Domain adaptation based on adjustable discriminator

作　　者：赵小强[1,2,3] 蒋红梅[1] ZHAO Xiao-qiang;JIANG Hong-mei(College of Electrical Engineering and Information Engineering, Lanzhou Univ. of Tech., Lanzhou 730050, China;Key Laboratory of Gansu Advanced Control for Industrial Processes, Lanzhou Univ. of Tech., Lanzhou 730050, China;National Experimental Teaching Center of Electrical and Control Engineering, Lanzhou Univ. of Tech., Lanzhou 730050, China)

机构地区：[1]兰州理工大学电气工程与信息工程学院,甘肃兰州730050 [2]兰州理工大学甘肃省工业过程先进控制重点实验室,甘肃兰州730050 [3]兰州理工大学国家级电气与控制工程实验教学中心,甘肃兰州730050

出　　处：《兰州理工大学学报》2022年第1期85-90,共6页Journal of Lanzhou University of Technology

基　　金：国家自然科学基金(62163023);国家重点研发计划项目(2020YFB1713600);甘肃省教育厅产业支撑计划项目(2021CYZC-02)。

摘　　要：当前基于对抗学习的领域适应(DA)对目标样本的适应性较差,导致目标域的预测精确度较低,为此提出基于可调节判别器的领域适应(A-DADA)算法.首先,利用两个判别器分类概率的距离作为权重应用到目标域对抗训练损失函数中,旨在减少已对齐目标样本对抗训练的次数同时增加未对齐目标样本的对抗训练次数;其次,将平方熵损失函数作为最小熵损失函数,提高类平衡性;最后,使用Office-31数据集将该算法与JAN等算法进行对比实验,实验结果表明,与JAN算法相比,该算法的平均精度值提高了2.5%.The current domain adaptation(DA)based on adversarial learning has poor adaptability to target samples,lower prediction accuracy of the target domain is caused.For this reason,this paper proposes a domain adaptation(A-DADA)algorithm based on an adjustable discriminator.First,the distance between the classification probabilities of the two discriminators is used as the weight to apply to the target domain confrontation training loss function,aiming to reduce the number of confrontation training times for aligned target samples and increase the number of confrontation training times for unaligned target samples.Secondly,square entropy loss function is used as the minimum entropy loss function to improve the class balance.Finally,the Office-31 dataset is used to compare the algorithm with JAN and other algorithms,the experimental results show that compared with the JAN algorithm,the average accuracy of the algorithm is increased by 2.5%.

关 键 词：领域适应 可调节判别器 对抗训练 平方熵损失 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]