Knowledge Distillation Model for Resource-Constrained Iris Segmentation in Biometric Systems

The country requires a digital identification system that is reliable, secure, efficient, and resistant to forgery. Iris biometric systems offer a promising solution due to the unique and lifelong stable patterns of the iris, which are difficult to replicate. The use of deep learning technology has improved the accuracy of iris biometric systems, but often results in models with high computational demands. The knowledge distillation approach provides a solution by reducing model complexity without significantly compromising performance. This study utilizes DeepLabV3 as the teacher model and Lightweight U-Net as the student model for iris image segmentation. The evaluation focuses on segmentation performance using IoU, precision, recall, and F1-score metrics, as well as model efficiency based on parameter count, model size, and inference time. The experimental results show that the student model achieves an IoU of 0.8823, precision of 0.9691, recall of 0.9078, and F1-score of 0.9374. Efficiency evaluation shows that the student model contains only 118,913 parameters, has a model size of 0.48 MB, and an inference time of 1.17 milliseconds. A comparison with the teacher model reveals that knowledge distillation reduces the number of parameters by 99.7 percent, the model size by 99.68 percent, and accelerates inference time by 50.21 percent. The reduction in segmentation accuracy is relatively small, with a 2.97 percent decrease in IoU, 0.41 percent in precision, 2.67 percent in recall, and 1.59 percent in F1-score. These findings confirm that knowledge distillation is an effective approach for producing lightweight and efficient segmentation models suitable for deployment on devices with limited computational resources.