A Robust Domain Adaptation Method Based on Class-Level Alignment and Gradient Penalty

Abstract

This study proposes an improved DANN domain adaptation algorithm to improve the generalization ability of the model in cross-domain classification tasks. Traditional DANN narrows the distribution difference between the source and target domains through adversarial training, but it still has limitations in class-level alignment, training stability, and feature expression ability. To this end, we introduce class-wise alignment to ensure that features of different categories match between the source and target domains, thereby reducing category confusion. In addition, the gradient penalty mechanism is used to enhance the stability of adversarial training and avoid gradient vanishing and mode collapse problems. At the same time, we design an adaptive feature enhancement (AFE) module, which combines multi-scale feature extraction and attention mechanism to improve the feature expression ability of target domain data. Experiments are evaluated on the Office-Home dataset. The results show that the improved DANN method outperforms the traditional DANN in all cross-domain tasks, especially when the target domain data is small, it can still maintain a high classification accuracy. This study provides a more stable and effective optimization strategy for domain adaptation methods, and provides new ideas for cross-domain learning tasks in computer vision, autonomous driving, medical image analysis and other fields.