A Hybrid Graph Neural Network Framework Integrating Handcrafted Features for Real-Time Iris Recognition in Motion

• Usman A. ABDURRAHMAN

  Department of Information and Communication Technology, Northwest University, Kano, Kano State, Nigeria

  Author

• Abdulkadir A. BICHI

  Department of Software Engineering, Northwest University, Kano, Kano State, Nigeria

  Author

• Usman HARUNA

  Department of Software Engineering, Northwest University, Kano, Kano State, Nigeria

  Author

• Akibu M. ABDULLAHI

  School of Computing and Informatics, Albukhary International University, Alor Setar, Malaysia

  Author

Iris recognition in dynamic environments, such as surveillance footage or real-time video streams, remains a significant challenge due to motion blur, occlusion, and the high computational cost of processing sequential frames. While traditional texture-based methods like Gabor filters struggle with motion deformations, modern deep learning approaches, particularly Graph Neural Networks (GNNs), offer superior spatial analysis but often at the expense of real-time performance. This paper proposes a novel hybrid framework that addresses these limitations by integrating handcrafted feature descriptors directly into a GNN architecture. Rather than relying solely on learned embeddings, the model initializes graph nodes using a fusion of traditional texture patterns and deep features, providing a richer and more resilient representation of the iris structure from the outset. Furthermore, we introduce a lightweight message-passing mechanism optimized for edge deployment, significantly reducing latency to meet the 25–30 frames-per-second requirement of real-time systems. By combining the interpretability and speed of traditional methods with the adaptive power of graph-based deep learning, the proposed approach enhances recognition accuracy under motion conditions while ensuring scalability for large databases. The proposed method achieves 95.1% recognition accuracy across three datasets, improving upon the original GNN framework by 2.3 percentage points and traditional methods by 9.2 points. It maintains robust performance under motion blur at 89.2% accuracy while operating at 35 frames per second on edge hardware. The system scales efficiently to one million users with query times of just 5.3 milliseconds. Experimental results demonstrate that this hybrid strategy outperforms both standalone deep networks and conventional algorithms, offering a viable path toward practical, real-world iris recognition in motion.

• FJET_21_13_13

Copyright (c) 2026 FUDMA Journal of Engineering and Technology

This work is licensed under a Creative Commons Attribution-NonCommercial 4.0 International License.