Knowledge Distillation Using Entropy-Guided Dynamic Parameter Transfer for Wound Image Classification

This study proposes a novel model compression approach for wound image classification using Knowledge Distillation (KD) enhanced by Entropy-Guided Adaptive Scaling (EGAS). EGAS comprises Entropy-Guided Adaptive Temperature Scaling (EGATS) and Entropy-Guided Adaptive Alpha Scaling (EGAAS), which dynamically adjust temperature and loss balancing based on entropy. Unlike traditional KD with fixed parameters, EGAS ensures smoother knowledge transfer for uncertain cases while preserving sharp distinctions for confident predictions. From Table VIII experimental results show that EGAS significantly improves classification accuracy, enabling the deployment of lightweight, high-performance models in resource-constrained environments for real-time medical applications.