Logistic regression projection-based feature representation for visual domain adaptation

The performance of visual image recognizers is considerably degraded while the training and test image sets not to follow the same distribution. In this study, we propose a novel method for unsupervised domain adaptation, called logistic regression projection-based feature representation. The proposed method performs the semi-supervised learning method on both the source and the target domains to predict the pseudo-label values for unlabeled target data. We incorporate the predicted target data with the source training dataset in order to learn feature representation which can be compensated for the distribution mismatch between source and target data. The results of the experiments on adaptation to different visual domains demonstrated that this method can achieve superior classification accuracy compared to the state-of-the-art methods. Based on the quantitative evaluation, the proposed unsupervised domain adaptation method can reduce the error rates by 15.85% compared to a corresponding 1-nearest neighbor classifier.