PISLNet : Parallel Iris Segmentation and Localization Network

Abstract Iris recognition in non‐cooperative environments has attracted significant attention. While recent research primarily focuses on improving accuracy and robustness, achieving real‐time performance remains a key requirement for practical applications. In conventional iris recognition systems, performing iris segmentation followed by circle localization sequentially reduces overall real‐time efficiency. To concurrently enhance accuracy and speed, this paper presents a Parallel Iris Segmentation and Localization Network (PISLNet), which adopts a dual‐task parallel processing framework with a lightweight backbone for fast iris segmentation and localization. PISLNet directly produces iris localization and segmentation results in an end‐to‐end manner, eliminating post‐processing costs. Furthermore, by considering the pairwise relationship and size correlation between the iris inner and outer circles, we introduce a double‐circle‐pair loss (DCP Loss) to address optimization challenges for circular bounding boxes represented in polar coordinates. Extensive experiments on multiple iris segmentation and localization datasets demonstrate that PISLNet attains superior performance on both tasks, achieving approximately 1% improvement over compared methods on near‐infrared datasets and about 3% improvement on visible‐light datasets. The overall system latency decreases by around 10 ms on near‐infrared datasets and processing speed doubles on visible‐light datasets. These results indicate PISLNet's suitability for real‐world real‐time iris recognition applications and support deployment on resource‐constrained embedded devices. © 2026 Institute of Electrical Engineers of Japan and Wiley Periodicals LLC.