The reliability of sclera-based biometric systems directly depends on the precise segmentation of ocular regions. Although deep segmentation models achieve high accuracy in this task under controlled conditions, their practical applicability is often limited by the domain gap between available laboratory-acquired training data and more complex real-world conditions. One of the main causes of lower segmentation robustness on real-world data is partial coverage, or occlusion, of the eye. In this work, we address this issue through generative data augmentation. By leveraging diffusion models, we inpaint photorealistic occlusions into high-quality laboratory ocular images from the \mbox{SBVPI} dataset and utilize the augmented dataset to train segmentation models. Evaluation on the challenging MOBIUS dataset demonstrates that inpainting synthetic occlusions serves as an effective regularization method. The segmentation models successfully transfer the acquired knowledge to real-world instances of occlusions, achieving higher accuracy and robustness compared to the baseline segmentation model trained exclusively on original images without generated occlusions.
|
