Learning domain-invariant representation for generalizable iris segmentation

Abstract

Cross-domain iris segmentation (CDIS) seeks to transfer knowledge from a labeled source dataset to an unlabeled target dataset. Existing CNN-based iris segmentation methods commonly assume that training and application stages share the same data distribution and modality setting, thus their performance may decline substantially on open-domain iris datasets unseen before. Furthermore, the process of annotating pixel-wise labels is labor-intensive and time-consuming, resulting in limited applicability of these methods in realistic scenarios. Therefore, we propose a generic domain adaptation iris segmentation framework (DAIrisSeg), which can be flexibly incorporated into existing methods. First, a domain-sensitive feature whitening strategy is proposed to effectively mitigate the domain-specific styles while preserving the domain-invariant content, thereby improving the model’s generalizability to unknown domain distribution. We then utilize the prototype estimation and the context-similarity learning adapter to produce reliable segmentation labels. In addition, DAIrisSeg incorporates prior constraints of the iris to further refine the segmentation results. Extensive experiments on three iris datasets demonstrate that the proposed method has shown consistent improvements over state-of-the-art (SOTA) methods.