Please wait a minute...
浙江大学学报(工学版)
浙江大学学报(工学版)  2022, Vol. 56 Issue (4): 754-763    DOI: 10.3785/j.issn.1008-973X.2022.04.015
计算机技术、信息工程     
基于自监督任务的多源无监督域适应法
吴兰1(),王涵1,李斌全1,李崇阳1,孔凡士2
1. 河南工业大学 电气工程学院，河南 郑州 450001
2. 郑州铁路职业技术学院，河南 郑州 450001
Multi-source unsupervised domain adaption method based on self-supervised task
Lan WU1(),Han WANG1,Bin-quan LI1,Chong-yang LI1,Fan-shi KONG2
1. School of Electrical Engineering, Henan University of Technology, Zhengzhou 450001, China
2. Zhengzhou Railway Vocational and Technical College, Zhengzhou 450001, China
 全文: PDF(1267 KB)   HTML
摘要：

针对多源聚合下同时对齐域不变特征较困难而造成分类精度不高的问题, 提出基于自监督任务的多源无监督域适应法. 该方法引入旋转、水平翻转和位置预测3个自监督辅助任务, 通过伪标签性、语义信息的一致性对无标签数据进行自适应的对齐优化. 构建新的优化损失函数, 减少多域公共类别的分类差异. 针对类别不均衡的问题, 基于少样本大权重的原则, 定义动态权重参数, 提高模型的分类性能. 在公开的Office-31、Office-Caltech10 2种基准数据集上, 与现有的主流方法进行实验对比. 实验结果表明, 在类别均衡、不均衡2种情况下, 分类精度最高可以提高6.8%.
关键词： 自监督任务;  类别不均衡;  语义信息;  权重;  域自适应    
Abstract:

A multi-source unsupervised domain adaptation method based on self-supervised tasks was proposed aiming at the problem of low classification accuracy due to the difficulty of simultaneously aligning domain-invariant features under multi-source aggregation. The method introduced three self-supervised auxiliary tasks of rotation, horizontal flip and position prediction, and performed adaption alignment optimization on unlabeled data through pseudo-labeling and consistency of semantic information. A new optimized loss was built, and the classification variance of multi-domain common classes was reduced. Dynamic weight parameters were defined to improve the classification performance of the model based on the principle of few samples and large weights for the problem of class-imbalance. Experiments were compared with the existing mainstream methods on the two benchmark data sets, Office-31 and Office-Caltech10. The experimental results show that the classification accuracy can be improved by up to 6.8% in the two cases of class balance and imbalance.
Key words: self-supervised task    class-imbalance    semantic information    weight    domain adaptation
收稿日期: 2021-06-01 出版日期: 2022-04-24
CLC:  TP 391  
基金资助: 国家自然科学基金资助项目（61973103）；河南省优秀青年科学基金资助项目；郑州市协同创新专项资助项目（21ZZXTCX01）
作者简介: 吴兰（1981—），女，教授，博士，从事深度学习的研究. orcid.org/0000-0002-2497-6556. E-mail： wulan@haut.edu.cn
服务　　
把本文推荐给朋友
加入引用管理器
E-mail Alert
作者相关文章　　
吴兰
王涵
李斌全
李崇阳
孔凡士

引用本文:

吴兰,王涵,李斌全,李崇阳,孔凡士. 基于自监督任务的多源无监督域适应法[J]. 浙江大学学报(工学版), 2022, 56(4): 754-763.

Lan WU,Han WANG,Bin-quan LI,Chong-yang LI,Fan-shi KONG. Multi-source unsupervised domain adaption method based on self-supervised task. Journal of ZheJiang University (Engineering Science), 2022, 56(4): 754-763.

链接本文:

https://www.zjujournals.com/eng/CN/10.3785/j.issn.1008-973X.2022.04.015    或    https://www.zjujournals.com/eng/CN/Y2022/V56/I4/754

图 1  自监督任务的对齐效果图
图 2  网络结构图
图 3  自监督任务语义一致原理
图 4  类别不均衡下的迁移原理
图 5  总体损失函数框架图
图 6  2种数据集的示例图
标准 方法 A/%
AW-D AD-W DW-A 平均值
单源迁移 DAN[2] 99.0 96.0 54.0 83.0
单源迁移 ADDA[3] 99.4 95.3 54.6 83.1
单源迁移 RevGrad[4] 99.2 96.4 53.4 83.0
源域组合 DAN[2] 98.8 96.2 54.9 83.3
源域组合 JAN[5] 99.4 95.9 54.6 83.3
源域组合 MCD[6] 99.5 96.2 54.4 83.4
源域组合 RevGrad[4] 98.8 96.2 54.6 83.2
多源迁移 DCTN[7] 99.6 96.9 54.9 83.8
多源迁移 M3SDA[10] 99.4 96.2 55.4 83.7
多源迁移 MDAN[8] 99.2 95.4 55.2 83.3
多源迁移 MDDA[11] 99.2 97.1 56.2 84.2
多源迁移 本文方法 99.6 97.2 56.4 84.4
表 1  Office-31数据集上的准确率比较分析
标准 方法 A/%
ADC-W ACW-D AWD-C DWC-A 平均值
源域组合 Source only 99.0 98.3 87.8 86.1 92.8
源域组合 DAN[2] 99.3 98.2 89.7 94.8 95.5
多源迁移 Source only 99.1 98.2 85.4 88.7 92.9
多源迁移 FADA[13] 88.1 87.1 88.7 84.2 87.1
多源迁移 DAN[2] 99.5 99.1 89.2 91.6 94.9
多源迁移 DCTN[7] 99.4 99.0 90.2 92.7 95.3
多源迁移 M3SDA[10] 99.4 99.2 91.5 94.1 96.1
多源迁移 本文方法 99.5 99.2 90.2 93.3 95.5
表 2  Office-Caltech10数据集上的准确率比较分析
图 7  Office-31数据集中损失函数对模型性能的影响分析
图 8  Office-Caltech10数据集中损失函数对模型性能的影响
图 9  Office-31数据集中对自监督任务训练效果的分析
图 10  域间差异损失中各损失函数对分类精度影响
图 11  Office-31数据集中各源域间类别样本量的对比
图 12  Office-Caltech10数据集中各源域间类别样本量的对比
图 13  Office-31各样本权重设置图示
图 14  Office-Caltech10各样本权重设置图示
类型 A/%
AD-W AW-D DW-A 平均值
类别不均衡 81.0 82.6 42.4 68.7
类别不均衡（权重） 87.8 85.5 42.7 72.0
表 3  不均衡样本下Office-31数据集上的精度对比分析
类型 A/%
ADC-W DWC-A AWD-C ACW-D 平均值
类别不均衡 75.3 67.8 63.7 79.3 71.5
类别不均衡（权重） 83.5 76.4 67.3 83.0 77.6
表 4  不均衡样本下Office-Caltech10数据集上的精度对比分析
1 沈宗礼, 余建波 基于迁移学习与深度森林的晶圆图缺陷识别[J]. 浙江大学学报: 工学版, 2020, 54 (6): 1228- 1239
SHEN Zong-li, YU Jian-bo Wafer map defect recognition based on transfer learning and deep forest[J]. Journal of Zhejiang University: Engineering Science, 2020, 54 (6): 1228- 1239
2 LONG M, CAO Y, WANG J, et al. Learning transferable features with deep adaptation networks [C]// International Conference on Machine Learning. Lille: [s. n.], 2015: 97-105.
3 TZENG E, HOFFMAN J, SAENKO K, et al. Adversarial discriminative domain adaptation[C]// Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. Honolulu: IEEE, 2017: 7167–7176.
4 GANIN Y, LEMPITSKY V. Unsupervised domain adaptation by backpropagation [C]// International Conference on Machine Learning. Lille: PMLR, 2015: 1180–1189.
5 LONG M, ZHU H, WANG J, et al. Deep transfer learning with joint adaptation networks[C]// International Conference on Machine Learning. Sydney: PMLR, 2017: 2208–2217.
6 SAITO K, WATANABE K, USHIKU Y, et al. Maximum classifier discrepancy for unsupervi-sed domain adaptation [C]//Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. Salt Lake City: IEEE, 2018: 3723–3732.
7 XU R, CHEN Z, ZUO W, et al. Deep cockt-ail network: multi-source unsupervised domain adaptation with category shift [C]//Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. Salt Lake City: IEEE, 2018: 3964–3973.
8 ZHAO H, ZHANG S, WU G. Adversarial multiple source domain adaptation [C]// Proceedings of the 32nd International Conference on Neural Information Processing Systems. Montreal: MIT Press, 2018: 8568–8579.
9 李威, 王蒙. 基于渐进多源域迁移的无监督跨域目标检测[EB/OL]. （2020-03-20）. http://kns.cnki.net/kcms/detail/11.2109.TP.20200320.1044.003.html.
LI Wei, WANG Meng. Unsupervised cross-domain object detection based on progressive multi-source transfer [EB/OL]. (2020-03-20). http://kns.cnki.net/kcms/detail/11.2109.TP.20200320.1044.003.html.
10 PENG X, BAI Q, XIA X, et al. Moment matching for multi-source domain adaptation [C]// Proceedings of the IEEE/CVF International Conference on Computer Vision. Seoul: IEEE, 2019: 1406–1415.
11 ZHAO S, WANG G, ZHANG S, et al. Multi-source distilling domain adaptation [C]// Proceedings of the AAAI Conference on Artificial Intelligence. New York: AAAI, 2020: 1297–1298.
12 ZHU Y, ZHUANG F, WANG D. Aligning domain specific distribution and classifier for cross-domain classification from multiple sources [C]// Proceedings of the AAAI Conference on Artificial Intelligence. Hawaii: AAAI, 2019: 5989–5996.
13 PENG X, HUANG Z, ZHU Y, et al. Federatedadversarial domain adaptation [C]// International Conference on Learning Representations. Addis Ababa: [s. n.], 2020.
14 ZHANG R, ISOLA P, ALEXEI A. Colorful image colorization [C]// European Conference on Computer Vision. Cham: Springer, 2016: 649–666.
15 GUSTAV L A, MICHAEL M, GREG O, et al. Learning representations for automatic colorization [C]// European Conference on Computer Vision. Amsterdam: [s. n.], 2016: 577–593.
16 CARL V, ABHINAV S, ALIREZA F, et al. Tracking emerges by colorizing videos [C]// Proceedings of the European Conference on Computer Vision. Munich: [s. n.], 2018: 391–408.
17 NOROOZI M, FAVARO P. Unsupervised learning of visual representations by solving jigsaw puzzles [C]// European Conference on Computer Vision. Amsterdam: [s. n.], 2016: 69–84.
18 CARL D, ABHINAV G, ALEXEI A. Unsupervised visual representation learning by conte-xt prediction [C]// Proceedings of the IEEE International Conference on Computer Vision. [S. l.]: IEEE, 2015: 1422–1430.
19 IMON J, PAOLO F. Self-supervised feature learning by learning to spot artifacts [C]// Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. Salt Lake City: IEEE, 2018: 2733–2742.
20 SPYROS G, PRAVEER S, NIKOS K. Unsupervised representation learning by predicting image rotations [EB/OL].(2018-03-21). https://doi.org/10.48550/arXiv.1803.07728.
21 DEEPAK P, PHILIPP K, JEFF D, et al. Context encoders: feature learning by inpainting [C]// Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. Las Vegas: IEEE, 2016: 2536–2544.
[1] 黄新宇,游帆,张沛,张昭,张柏礼,吕建华,徐立臻. 基于多分类及特征融合的静默活体检测算法[J]. 浙江大学学报(工学版), 2022, 56(2): 263-270.
[2] 黎振宇,陈晓国,宋永超,龚建平,余志纬,朱永兴,鲁周勋. 二元联系数-投影灰靶决策理论在电网应急能力评估中的应用[J]. 浙江大学学报(工学版), 2021, 55(5): 927-934.
[3] 王友卫,凤丽洲. 基于合群度-隶属度噪声检测及动态特征选择的改进AdaBoost算法[J]. 浙江大学学报(工学版), 2021, 55(2): 367-376.
[4] 赵小虎,尹良飞,赵成龙. 基于全局?局部特征和自适应注意力机制的图像语义描述算法[J]. 浙江大学学报(工学版), 2020, 54(1): 126-134.
[5] 郭洪武,蒲雷,张予燮,吴静,赵蕊,谭忠富. 基于粗糙集理论的风光蓄互补系统优化模型[J]. 浙江大学学报(工学版), 2019, 53(4): 801-810.
[6] 董月,冯华君,徐之海,陈跃庭,李奇. Attention Res-Unet: 一种高效阴影检测算法[J]. 浙江大学学报(工学版), 2019, 53(2): 373-381.
[7] 许越, 徐之海, 冯华君, 李奇, 陈跃庭, 徐毅, 赵洪波. 双场景类型遥感图像的配准拼接优化[J]. 浙江大学学报(工学版), 2019, 53(1): 107-114.
[8] 谢宪毅, 金立生, 高琳琳, 夏海鹏. 基于变权重系数的LQR车辆后轮主动转向控制研究[J]. 浙江大学学报(工学版), 2018, 52(3): 446-452.
[9] 王昱, 傅莉, 梁宵, 李勇. 混合缺失信息集结下的属性赋权方法[J]. 浙江大学学报(工学版), 2016, 50(9): 1806-1814.
[10] 王仁超, 邓剑, 曹婷婷. 基于改进VIKOR法的进度计划方案选择问题研究[J]. 浙江大学学报(工学版), 2016, 50(10): 2025-2030.
[11] 王思照, 张仪萍. 基于T5单元的体积不可压缩问题光滑有限元法[J]. 浙江大学学报(工学版), 2015, 49(10): 1967-1973.
[12] 寿黎但, 廖定柏, 徐昶, 陈刚. PWLRU: 一种面向闪存数据库的缓冲区存取算法[J]. J4, 2010, 44(12): 2257-2262.
[13] 武守飞, 王正肖, 潘晓弘, 等. 基于粗集理论的产品属性定制权重确定方法[J]. J4, 2009, 43(12): 2250-2253.
[14] 郭勇 孙炳楠 叶尹. 多目标优化方法在输电塔阻尼器布置中的应用[J]. J4, 2006, 40(10): 1755-1760.