ISL-Net: dual-stream interaction network with task-optimized modules for more accurate, complete iris segmentation and localization

Abstract

Iris images captured in uncooperative and unconstrained environments pose significant challenges for iris segmentation and localization owing to factors including high occlusions, specular reflections, motion blur, iris rotation, and off-angle images. To address this challenge, this paper proposes ISL-Net, a multitask segmentation network with a task-optimization module based on deep learning for joint iris segmentation and localization. We developed a dual-stream interactive module (DSIM) that combines dual-stream decoders to facilitate information exchange between tasks without interference. To optimize the iris-segmentation and iris-localization performance, we incorporated a balanced attention module (BAM) and a boundary-enhancement module (BEM) in the skip connections of the respective task stream decoders. The BEM recovers missing boundaries in iris localization, while the BAM focuses on uncertain areas in iris segmentation, enhancing the model’s ability to handle these regions. These modules complement each other, improving overall system performance without interference. The proposed model was evaluated on three challenging iris datasets, outperforming most existing models by achieving e1 index scores of 0.34, 0.79, and 0.61% and average normalized Hausdorff distances (HDs) of 0.7221, 1.1914, and 1.0396%. The results indicate that ISL-Net can generate normalized iris images with simple post-processing, making it suitable for direct application in existing iris-recognition systems.

Graphical Abstract