Comparative study on visual detection methods of drugs based on YOLO model

The accuracy and real-time performance of pill sorting systems are critical for medical safety. While YOLO series models are efficient, traditional methods struggle with small pill targets and complex backgrounds. This study trained YOLOv5/v8/v11 on a self-built dataset (261 categories, 4,000 images), analyzing mAP, FPS, and model complexity. The proposed CSFF module reduced small-target missed detection by 18%, while revealing version-specific applicability: YOLOv8 balances accuracy/speed, and YOLOv11 suits low-power devices. Iterative updates have progressively enhanced performance through feature pyramid optimization and adaptive anchor adjustments. These findings provide theoretical guidance for medical automation scenarios.