An efficient iris segmentation paradigm for low-quality eye movement images

Abstract Iris segmentation is widely used in medical imaging, wearable devices, and security applications. However, under resource-constrained conditions, existing iris segmentation models often face efficiency challenges and distribution shifts arising from low-quality images. An efficient iris segmentation paradigm is proposed in this study to address these two challenges. The proposed paradigm comprises an image preprocessing strategy, a novel iris segmentation model (ISNet), and an iris contour processing (ICP) module. The ISNet enhances segmentation accuracy through its innovative Deformable Multi-Receptive Field Residual Attention module, which integrates deformable convolutions, multiscale dilated blocks, and residual attention mechanisms within a U-Net++ framework. The ICP module is employed to optimize the smoothness of the iris edges. The effectiveness of this paradigm was verified through comparative experiments on the TEyeD dataset and a self-collected low-quality iris dataset (LIrisD), image degradation experiments, and application experiments on the RK3588 edge device. Specifically, we obtain 93.5%/92.5% mean intersection over union on TEyeD/LIrisD in non-deployment (desktop) evaluation and 85.5%/82.5% on the RK3588 edge device, demonstrating high accuracy under acquisition-device constraints and superiority over existing iris segmentation methods.