RDAOT: Robust Unsupervised Deep Sub-Domain Adaptation Through Optimal Transport for Image Classification

In traditional machine learning, the training and testing data are assumed to come from the same independent and identical distributions. This assumption, however, does not hold up in real-world applications, as differences between the training and testing data may have different distributions. Domain adaptation has emerged as a solution that enables the transfer of knowledge between domains with distinct distributions. In this paper, we primarily utilize domain adaptation in the context of visual recognition tasks despite its growing application in diverse domains. Earlier studies have mainly aimed at minimizing the differences in global distributions between the domains and failed to capture the local, pertinent features crucial for domain alignment. Furthermore, models struggle to perform well and generalize to target data when outliers or noise exist in the datasets. This work addresses these problems and provides unique strategies for unsupervised domain adaptation using RDAOT (Robust Deep Adaptation via Optimal Transport). To capture local information by utilizing LMMD (Local Maximum Mean Discrepancy) to minimize the divergence of the feature distributions between the domains. We examine label noise robustness in the source domain and ROT (Robust Optimal Transport) loss to preserve robustness in domain adaptation, which lessens the cost of transporting source distributions to the target distributions. The significance of our presented technique was assessed through extensive experiments on six different visual recognition domain adaptation datasets. The results demonstrate that our method outperforms the current state-of-the-art techniques, indicating superior performance. Our approach was evaluated against several baselines, and the results significantly improved average accuracy across various datasets. Specifically, the average accuracy improved from on the OfficeCaltech10 (91.8% to 96.85%), OfficeHome (67.7% to 68.10%), Office31 (88.17% to 88.92%), IMAGECLEF-DA (87.9% to 90.24%), PACS (69.08% to 85.72%), and VisDA-2017 (80.2 % to 89.43%) datasets, respectively.