A Domain Adaptive Adversarial Training Method Based on Self-Supervised Learning

Abstract

Image classification technology based on neural network is an important task in computer vision, and the introduction of transfer learning can solve the problems of lack of data sets and long training time. To address this problem, this paper proposes a self-supervised domain-adaptive adversarial network approach. The algorithm uses the VGG network to extract image features, realizes the transfer learning of different image styles through domain adversarial training, and introduces a data augmentation model and self-supervised learning method based on pseudo-label to improve the accuracy of model classification. The experimental results show that the model can effectively improve the accuracy of image transfer learning of different styles in the image classification problem. When the number of pseudo-labels is 10, the classification effect is the best, and the accuracy rate is improved to 12.99%, which greatly saves training time and computing power while solving the problem of missing training data.