Semi-supervised domain generalization with clustering and contrastive learning combined mechanism

Domain generalization (DG) is essential in numerous tasks and strives to obtain a well-generalized model under multi-source data distributions. However, achieving the requirement that all data have precise annotations in real-life deployment poses a significant challenge. Therefore, we turn to semi-supervised domain generalization (SSDG), which relaxes this requirement from full annotation to limited annotation. The task presents two challenging issues: (1) learning pseudo-labels for unlabeled data under multi-source data distributions, and (2) effectively utilizing a minority of labeled data alongside a majority of unlabeled data to extract robust feature representations. To address these challenges, we propose an innovative approach that employs a collaborative mechanism of clustering and contrastive learning (CMCC) for high-quality pseudo-labels. Specifically, we use clustering to select high-confidence unlabeled samples and subsequently enhance the reliability of their pseudo-labels using contrastive learning. Furthermore, we introduce discrete Fourier transform (DFT)-based consistency learning (DCL) to enable the model to focus on domain-invariant features hidden in the phase information by mixing the amplitude spectra of two random images, thereby mitigating the domain shift issue. Comprehensive experiments conducted on two standard DG benchmark datasets with limited annotations validate the effectiveness of the proposed approach.