Abstract
Automated identification of pharmaceutical drugs and vitamins is critical for improving medication safety and reducing human errors in healthcare systems. This study proposes a novel multi-scale, attention-enhanced convolutional neural network designed for fine-grained classification of pharmaceutical images. The model integrates multi-scale feature extraction to capture both local visual details (e.g., imprints, edges) and global characteristics (e.g., shape, color distribution), while an attention mechanism emphasizes discriminative regions, enabling accurate differentiation of visually similar classes. Transfer learning with ImageNet-pretrained weights accelerates convergence and improves generalization. The model was evaluated on the Pharmaceutical Drugs and Vitamins Synthetic Images dataset, comprising 100,000 images across ten classes. Experimental results demonstrate high classification performance, achieving a macro-averaged F1-score of 0.897, with per-class F1-scores ranging from 0.766 to 0.934. ROC analysis and confusion matrices confirm robust discriminative capability, although challenges remain for classes with minimal visual differences. Compared to conventional CNN architectures, the proposed approach shows improved precision, recall, and F1-scores. The study contributes a scalable, accurate framework for automated pharmaceutical identification, offering a foundation for integrating synthetic data augmentation, multimodal features, and few-shot learning in future research.