SwinIrisNet: A Hybrid Deep Learning Framework for Robust Iris Segmentation

SwinIrisNet is a memory-efficient deep learning architecture for iris segmentation that combines a Swin Transformer encoder with a CNN residual encoder in a U-Net framework. The dual-stream encoder captures both global iris geometry (via windowed self-attention) and fine local boundary details (via convolutional residual blocks), fusing them through a cross-attention mechanism with CBAM at each encoder stage. The decoder uses ASPP for multi-scale context and attention-gated skip connections for precise boundary recovery. Training employs a boundary-aware composite loss (BCE + Dice + Sobel-edge boundary loss). This repository provides the complete source code, configuration, and SageMaker-compatible training notebooks for reproducing the ablation study and final model reported in the accompanying paper. Four model variants are included: (1) Swin-T encoder only, (2) CNN encoder only, (3) full hybrid with standard BCE+Dice loss, and (4) the final SwinIrisNet with boundary-aware loss. Key results (CASIA Thousand / UBIRIS.v2): F1 Score: 0.9605 / 0.9624mIoU: 0.920 / 0.930NICE2 (boundary error): 1.82% / 1.98%Parameters: 13.58M | FLOPs: 20G | Inference: 183 ms (CPU)Requirements: Python 3.10, TensorFlow 2.12. Training designed for AWS SageMaker (ml.g5.xlarge) but runnable locally.