Contrastive Domain Adaptation by Minimizing Divergence in Source-Target Image Distributions

Images acquired under varying imaging conditions (domains) can introduce shifts in data distributions, which negatively affects a classifier’s performance trained on data from a different source. Existing domain adaptation solutions use contrastive or adversarial learning without explicitly learning data distributions. This paper demonstrates a novel framework to learn source and target data distributions and minimize the distribution divergence jointly with contrastive image representation learning. The proposed method improves classification accuracy by $5.1 \%$ to $18.4 \%$ against two baselines methods without domain adaptation (cross-entropy only and cross-entropy with contrastive learning) and by $1.7 \%$ to $27.8 \%$ against stateof-the-art domain adaptation methods across six adaptation scenarios. With sufficient source data samples, self-supervised pretraining before domain adaptation yields the best accuracy in four of six scenarios. The proposed image distribution learning provides direct and explainable controls over domain adaptation where the distribution divergence loss jointly converges with contrastive image representation learning.