Year: 2026
Type: article
Abstract: Knowledge distillation (KD) facilitates the compression of large, high-performing neural networks into efficientstudent models, enabling deployment on resource-limited devices like mobile phones and IoT systems. This paperintroduces a KD methodology, which involves training a student to capture a teacher’s soft labels, intermediate featurerepresentations, and ground truth labels, ensuring both compactness and accuracy. Applied to face recognition, our hybridKD framework trains a MobileFaceNet student under an InceptionResNetV1 teacher, achieving 90.40% accuracy and a96.25% AUC, outperforming lightweight models while remaining suitable for edge devices. These results highlight thepotential of KD to enable robust, scalable face recognition solutions for real-world, resource-constrained environments.
Source: ENP Engineering Science Journal
Institution: Polytechnic School of Algiers
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: diamond
OpenAlex ID: W7123360366