Unified Iris Segmentation and Recognition with Hybrid Mask2Former-Swin Transformer Architecture for Biometric Authentication

Abstract

Although iris recognition is widely recognized as a highly dependable biometric modality, conventional systems usually segregate segmentation and classification into distinct steps, which results in error propagation and subpar overall performance. In this paper, we present an end-to-end Transformer-based architecture that simultaneously learns identity categorization and accurate iris segmentation within a single computational graph. Initially, a Mask2Former decoder based on a Swin-Tiny backbone creates high-fidelity binary iris masks by optimizing a combined Dice–Focal loss to improve boundary delineation and combining multi-scale data via cross-attention. After resizing the masked iris picture to 224 by 224 pixels, it is fed into a Swin-Base transformer, which extracts robust radial and circular texture representations via hierarchical shifted-window self-attention and patch-merging layers.These embeddings are mapped to C identity classes under a cross-entropy loss by a lightweight classification head that consists of two linear layers and a softmax activation. Our combined loss function guarantees mutual feedback between the two tasks by giving the segmentation and classification losses similar weights, which speeds up convergence and improves accuracy. The suggested method outperforms traditional two-stage pipelines, achieving average accuracy, recall, and F1-score above 98.5% and producing more compact, easily deployable models, according to extensive experiments conducted on three benchmark datasets (CASIA-Iris-Thousand, CASIA-Iris-Lamp, and CASIA-IntervalV4). The efficiency of integrated Transformer topologies for reliable, high-precision iris authentication is demonstrated in this work.