Polar Topology Transformers With Anatomical Skip Connections for Efficient Iris Segmentation

This paper proposes RadialSwin-UNet, a lightweight iris-specific segmentation framework that builds upon Swin-based modeling and incorporates topology-aligned polar representation for robust iris delineation under occlusion and illumination variation. Instead of operating purely in the Cartesian plane, our method adopts a compact polar representation that matches the iris' radial–angular geometry, allowing the model to capture global circular structure more naturally. A crop-based polar unwrapping focuses computation on the iris region and stabilizes radial–angular topology for efficient attention modeling, while anatomically informed priors encourage ring-consistent predictions under occlusion and illumination variation. Experiments on CASIA-IrisV4-Interval, CASIA-IrisV4-Lamp, and IITD show that the proposed method provides a favorable accuracy–efficiency trade-off, achieving a Dice score of 0.9885 on CASIA-IrisV4-Interval while using 13× fewer parameters than U-Net (2.4M vs. 32M), together with 3.4× faster training and 1.9–3.2× faster inference. These results suggest that the proposed method is a promising candidate for efficient biometric deployment under limited computational resources.