Wang Ziqian, Tao Qing, Zhong Zhijian, Yang Ming, Wang Xuecheng, Luo Xiaorong, Wu Zhenfeng
National Key Laboratory for the Modernization of Classical and Famous Prescriptions of Chinese Medicine, Jiangxi University of Chinese Medicine, 330004 Nanchang, China; Jiangxi Drug Inspection Center, 330000 Nanchang, China.
Computer Institute, Jiangxi University of Chinese Medicine, 330004 Nanchang, China.
Int J Pharm. 2025 Jun 10;678:125704. doi: 10.1016/j.ijpharm.2025.125704. Epub 2025 May 10.
This study presents a real-time system that integrates deep learning and machine vision for defect detection and coating thickness measurement of irregularly shaped film-coated tablets. To overcome the accuracy and speed limitations of the traditional YOLOv8 model on irregular shapes, we propose an enhanced YOLOv8n architecture incorporating a Mamba-Like Linear Attention (MLLA) mechanism. This modification significantly improves the model's ability to detect subtle defects with higher precision. The system captures real-time tablet images using industrial cameras, ensuring reliable and accurate defect identification. For coating thickness measurement, the system employs sub-pixel image processing techniques to precisely measure the Feret diameter of tablets, while weight analysis is integrated to assess coating uniformity. By establishing a strong linear correlation between tablet diameter and weight, the system enables accurate estimation of coating thickness. Experimental validation demonstrates a Root Mean Square Error of Prediction (RMSEP) of 2.2 mg, which ensures highly precise weight monitoring throughout the coating process. The proposed system achieves an overall classification accuracy of 91.87 % across eight types of coated tablets, confirming its robustness and effectiveness. This innovative solution offers pharmaceutical manufacturers a scalable and cost-efficient tool for real-time quality assessment of irregularly shaped tablets, enhancing production efficiency, optimizing quality control, and minimizing defects in continuous manufacturing environments.
本研究提出了一种实时系统，该系统集成了深度学习和机器视觉，用于对不规则形状的薄膜包衣片剂进行缺陷检测和包衣厚度测量。为了克服传统YOLOv8模型在处理不规则形状时的精度和速度限制，我们提出了一种增强的YOLOv8n架构，该架构融入了类似曼巴的线性注意力（MLLA）机制。这一改进显著提高了模型以更高精度检测细微缺陷的能力。该系统使用工业相机捕获片剂的实时图像，确保可靠且准确地识别缺陷。对于包衣厚度测量，系统采用亚像素图像处理技术精确测量片剂的费雷特直径，同时结合重量分析来评估包衣均匀性。通过建立片剂直径与重量之间的强线性相关性，该系统能够准确估计包衣厚度。实验验证表明预测均方根误差（RMSEP）为2.2毫克，这确保了在整个包衣过程中进行高精度的重量监测。所提出的系统在八种类型的包衣片剂上实现了91.87%的总体分类准确率，证实了其稳健性和有效性。这种创新解决方案为制药制造商提供了一种可扩展且经济高效的工具，用于对不规则形状片剂进行实时质量评估，提高生产效率、优化质量控制并在连续制造环境中最大限度地减少缺陷。
