A Knowledge Distillation-Based Lightweight Model for Efficient and Accurate Pill Classification

Resumen

Pills classification plays a crucial role in healthcare, ensuring medication safety and preventing errors in prescription dispensing. Traditional deep learning models have demonstrated promising results in pill recognition; however, they often require high computational resources, making them impractical for deployment in resource-constrained environments such as mobile healthcare applications and embedded systems. Existing approaches rely on large-scale deep learning models, which are difficult to deploy efficiently without compromising accuracy. To address this challenge, the proposed PillNet-KD (Pill Classification Network with Knowledge Distillation) employs a high-performing mentor model based on optimized Xception to transfer knowledge to a lightweight apprentice model, a Custom Convolutional Neural Network (CustomCNN), designed for efficient inference. Knowledge distillation ensures that the apprentice model captures essential feature representations from the mentor model through soft probability transfer. The PillNet-KD framework effectively combines high accuracy with reduced computational complexity. Extensive evaluation demonstrates that PillNet-KD achieves competitive performance with significantly lower inference time, making it suitable for real-time deployment. The framework offers a practical solution for enhancing automation in mobile health applications and intelligent medication dispensing systems.