EDIL-SegRayDP: Training-Free Iris Segmentation via Segmentation-First Ray-Wise Dynamic Programming

Iris segmentation remains a critical yet challenging stage in biometric recognition, especially under off- axis capture, eyelid and eyelash occlusion, specular reflections, and illumination variations that violate the circular and unobstructed assumptions of classical pipelines. We present EDIL-SegRayDP, a training-free and explainable iris segmentation framework that departs from the conventional localization-first paradigm by treating annulus recovery as the primary optimization objective. Rather than committing early to a global center/radius hypothesis and refining it afterward, the proposed method performs segmentation-first boundary recovery with segmentation-aware center rescue and fail-safe outer-boundary control. Occlusion is handled explicitly through geometry-normalized masking and validity-aware annulus construction, while all key parameters are defined in scale-normalized form for cross-dataset portability. Experiments under a fixed-configuration protocol on IITD and CASIA-IrisV4-Interval show strong non-CNN performance with CPU-only inference, achieving an iris-mask mean Dice of 0.9106 on IITD and 0.9377 on CASIA-IrisV4- Interval, with corresponding pupil Dice of 0.9763 and 0.9755. Additional full-benchmark evaluations on CASIA-IrisV4-Lamp and CASIA-IrisV4-Thousand further confirm the portability of the proposed framework across more challenging and larger-scale subsets. Under the evaluation protocol adopted in this study, these results compare favorably with a recent training-free reference, supporting EDIL-SegRayDP as a competitive and interpretable training-free alternative for iris segmentation under non-ideal imaging conditions.