Pill-MSVAE: A Multiscale Variational Autoencoder Based on Internet of Things for Accurate Automatic Drug Identification and Monitoring

Pill recognition is critical for preventing medication misuse and ensuring patient safety, yet it remains challenging due to the growing number of pharmaceutical products and their highly similar visual appearances. Existing pill recognition methods often rely on single-scale feature representations and deterministic models, which limits their robustness and generalization under complex imaging conditions such as blur, illumination variation, and pill wear. To address these limitations, we propose an IoT-based Multi-Scale Variational Autoencoder (Pill-MSVAE) for automatic pill image recognition. The proposed framework integrates convolutional neural networks and Transformers to jointly capture local texture details and global structural information, while variational encoding with reconstruction constraints enhances feature robustness and generalization. Extensive experiments on a public pill image dataset show that Pill-MSVAE achieves an F1-score of 0.9718, Accuracy of 0.9716, Precision of 0.9727, and AUC of 0.9989, consistently outperforming state-of-the-art methods across all evaluation metrics. These results validate the effectiveness of the proposed approach for reliable pill recognition and highlight its potential for practical deployment in IoT-enabled healthcare and medication management systems.