计算机工程与应用 ›› 2022, Vol. 58 ›› Issue (14): 153-159.DOI: 10.3778/j.issn.1002-8331.2101-0042

• 模式识别与人工智能 • 上一篇    下一篇

结合风格迁移的最小化类混淆领域自适应算法

梅校杰,张灵   

  1. 广东工业大学 计算机学院,广州 510006
  • 出版日期:2022-07-15 发布日期:2022-07-15

Domain Adaptive Algorithm for Minimizing Class Confusion Combined with Style Transfer

MEI Xiaojie, ZHANG Ling   

  1. School of Computer Science and Technology, Guangdong University of Technology, Guangzhou 510006, China
  • Online:2022-07-15 Published:2022-07-15

摘要: 在无监督领域自适应中分类器对目标域的样本进行类别预测时容易产生混淆预测,虽然已有研究提出了相关算法提取到样本的类间相关性,降低了分类器在目标域上的类混淆预测。但该方法仍然未能解决源域和目标域因共享特征稀疏导致的迁移学习能力不足的问题,针对这个问题,通过使用生成对抗网络对源域进行了风格迁移,扩展源域各类样本的特征空间可供目标域匹配的共享特征,解决因共享特征稀疏导致分类器正迁移力不足的问题,从而进一步减少分类器在目标域上产生的类混淆预测。当分类器利用扩充后的共享特征对目标域样本预测分类概率时,基于不确定性权重机制,加重预测概率权重使其能在几个预测概率峰值上以更高的概率值突出,准确地量化类混淆,最小化跨域的类混淆预测,抑制跨域的负迁移。在UDA场景下,对标准的数据集ImageCLEF-DA和Office-31的三个子数据集分别进行了领域自适应实验,相较于RADA算法平均识别精度分别提升了1.3个百分点和1.7个百分点。

关键词: 迁移学习, 最小化类混淆, 风格迁移, 领域自适应, 生成对抗网络

Abstract: In unsupervised domain adaptation, a classifier is easily to produce the confusion prediction when it predicts the class of samples of a target domain, though existing studies have proposed that the class correlation of samples is extracted by relevant algorithms which reduce a classifier’s confusion prediction in the target domain, this method remains problems unsolved that the transfer learning ability is inadequate in a source domain and target domain because of sparse shared features, targeted to which, through application of generative adversarial nets to migrate the style of a source domain, to extend shared features of various samples of feature space in source domain available for source domain matching to solve the problem of classifier’s insufficient positive transfer ability led by sparse shared features, thus to further reduce a classifier’s class confusion prediction produced in a target domain. When a classifier predicts the classified probability of samples of a target domain by using extended shared features, based on uncertainty of weighting system to add the weighting of predictive probability to make it be highlighted with higher probability values on several predicted probability peaks for an accurate quantitative class confusion, it minimizes cross-origin class confusion prediction and inhibits cross-origin negative transfer. Under UDA scene, domain adaptive experiments have been conducted respectively for three sub-data sets of standard data set ImageCLEF-DA and Office-31, and whose average recognition precision are higher by 1.3 percentage points and 1.7 percentage points respectively compared with RADA algorithm.

Key words: transfer learning, minimizing class confusion, style transfer, domain adaptation, generative adversarial network