Image classification based on unsupervised adversarial transfer learning and preserving the inter-class and intra-class distance

Document Type : Original Article


1 University of Tabriz

2 Faculty of Electrical and Computer Engineering, University of Tabriz

3 İstinye university


The paper explores the growing use of deep learning in machine vision, acknowledging challenges in model generalizability due to insufficient data. To address this issue, the proposed solution employs multi-source unsupervised adversarial transfer learning, enhancing adaptability across diverse datasets. This approach compels the network to learn shared features between different datasets rather than domain-specific ones. A novel loss function is introduced, emphasizing inter-class and intra-class distance preservation. This enhances the network's ability to learn similar representations for samples within the same class and dissimilar representations for instances across different classes. Evaluation involves testing on MNIST, MNIST-M, SVHN, and USPS datasets under various transmission scenarios. Comparative analysis with other algorithms demonstrates the effectiveness of the proposed approach, achieving accuracies of 99.5%, 98.8%, 98.5%, and 98.2% for MNIST, MNIST-M, SVHN, and USPS datasets, respectively. The results highlight the solution's success in addressing insufficient data challenges and improving model generalizability in machine vision applications


Main Subjects