Domain generalization for image classification with dynamic decision boundary

Abstract

Domain Generalization (DG) has been widely used in image classification tasks to effectively handle distribution shifts between source and target domains without accessing target domain data. Traditional DG methods typically rely on static models trained on the source domain for inference on unseen target domains, limiting their ability to fully leverage target domain characteristics. Test-Time Adaptation (TTA)-based DG methods improve generalization performance by adapting the model during inference using target domain samples. However, this often requires parameter fine-tuning on unseen target domains during inference, which may lead to forgetting of source domain knowledge or reduce real-time performance. To address this limitation, we propose a Dynamic Decision Boundary-based DG (DDB-DG) method for image classification, which effectively leverages target domain characteristics during inference without requiring additional training. In the proposed DDB-DG, we first introduce a Prototype-guide Multi-lever Prediction (PMP) module, which guides the dynamic adjustment of the decision boundary learned from the source domain by leveraging the correlation between test samples and prototypes. To enhance the accuracy of prototype computation, we also propose a data augmentation method called Uncertainty Style Mixture (USM), which expands the diversity of training samples to improve model generalization performance and enhance the accuracy of pseudo-labeling for target domain samples in prototypes. We validate DDB-DG using different backbone networks on three publicly available benchmark datasets: PACS, Office-Home, and VLCS. Experimental results demonstrate that our method achieves superior performance on both ResNet-18 and ResNet-50, surpassing the state-of-the-art DG and TTA methods.