A novel deep learning network for iris recognition using non-global interactive channel attention and lightweight spatial attention

Abstract

Iris recognition is acknowledged as a highly reliable and significant biometric recognition technology, with broad applications in diverse industrial sectors. Despite advancements, the performance of current iris recognition algorithms requires improvement, and their computational complexity necessitates further reduction. In response, we introduce a novel deep learning framework for iris recognition that employs an attention mechanism, leveraging ResNet-50 as the foundational architecture and integrating a lightweight attention module (LAM). Within the LAM, we propose a non-global interactive channel attention module (NGICAM) and a lightweight spatial attention module (LSAM), further optimizing the triplet loss to more effectively extract iris texture features. Our experimental results indicate that this approach yields a lower equal error rate (EER) and a higher decidability index (DI) compared to existing models, achieving the most significant EER reduction of 23.26% and an increase of up to 2.3405 in DI.