Pairwise Generalization Network for Cross-Domain Image Recognition
- PDF / 1,218,984 Bytes
- 19 Pages / 439.37 x 666.142 pts Page_size
- 80 Downloads / 221 Views
Pairwise Generalization Network for Cross-Domain Image Recognition Y. B. Liu1 · T. T. Han1 · Z. Gao1,2
© Springer Science+Business Media, LLC, part of Springer Nature 2019
Abstract In recent years, convolutional neural networks have received increasing attention from the computer vision and machine learning communities. Due to the differences in the distribution, tone and brightness of the training domain and test domain, researchers begin to focus on cross-domain image recognition. In this paper, we propose a Pairwise Generalization Network (PGN) for addressing the problem of cross-domain image recognition where Instance Normalization and Batch Normalization are added to enhance their abilities in the original domain and to expand to the new domain. Meanwhile, the Siamese architecture is utilized in the PGN to learn an embedding subspace that is discriminative, and map positive sample pairs aligned and negative sample pairs separated, which can work well even with only few labeled target data samples. We also add residual architecture and MMD loss for the PGN model to further improve its performance. Extensive experiments on two different public benchmarks show that our PGN solution significantly outperforms the state-of-the-art methods. Keywords Cross-domain · Image recognition · Pairwise
1 Introduction Deep convolutional neural networks (CNNs) have significantly improved the-state-of-thearts for diverse machine learning problems and applications, such as image recognition [1], object detection [2], and semantic segmentation [3]. Unfortunately, many of the existing
This work was supported in part by the National Natural Science Foundation of China (Nos. 61872270, 61572357, 61202168), Opening Foundation of Tianjin Key Laboratory of Intelligence Computing and Novel Software Technology, Tianjin University of Technology, China. Tianjin Municipal Natural Science Foundation (No. 18JCYBJC85500).
B
Z. Gao [email protected]
1
Tianjin Key Laboratory of Intelligence Computing and Novel Software Technology and Key Laboratory of Computer Vision and System, Ministry of Education, Tianjin University of Technology, Tianjin 300384, China
2
Qilu University of Technology (Shandong Academy of Sciences), Shandong Computer Science Center (National Supercomputer Center in Jinan), Shandong Artifical Intelligence Institute, Jinan 250014, People’s Republic of China
123
Y. B. Liu et al.
methods usually only apply to a particular domain and rely on data with a large number of labels. If the data of target domain is unavailable or expensive to label it, the conventional machine learning approach would drop markedly. To address the issue, one of the main methods is to use the transfer learning and domain adaptation to learn a model that is discriminative and domain-invariant. For domain adaptation, there are three categories: supervised domain adaptation (SDA) [4–6], unsupervised domain adaptation (UDA) [7,8], and semi-supervised domain adaptation [9,10]. UDA does not require target data to be labeled, but it expects large amoun
Data Loading...
