目的 为提升质检过程中药用空心胶囊的表面缺陷检测精度及其自动化水平。方法 通过设计高质量图像采集方案来避免胶囊表面出现光斑，以此构建药用空心胶囊缺陷数据集。基于YOLOv4算法，建立深度学习检测模型，利用多尺度特征提取以及训练策略，增强对小目标缺陷检测的鲁棒性。采用K-means++聚类算法更新锚框初始值，以提高模型对胶囊表面缺陷的预测性能。结果 实验结果表明，提出的胶囊缺陷检测方法能够准确判别胶囊好坏,并能检测出其表面的凹陷、孔洞、划痕、污点和接口缺损等5类缺陷，其中对于胶囊有无缺陷的平均精确均值达99.05%，各缺陷类型的平均精确率为91.81%，而每秒检测图像可达22张。与其他典型的目标检测方法相比，文中方法在检测速度和精度上都有一定优势。结论 文中所提出的基于YOLOv4的缺陷检测方法实现了对药用空心胶囊多类型缺陷的分类与定位，具有较好的检测效果和稳定性，在满足生产质量管控要求的同时，可大幅降低人工成本。

The work aims to improve the surface defect detection accuracy and automation level of the pharmaceutical hollow capsules in the process of quality inspection. A high-quality image acquisition scheme was designed to avoid light spots on the capsule surface. A dataset of defects in pharmaceutical hollow capsules was constructed. Based on the YOLOv4 algorithm, a deep learning detection model with multi-scale feature extraction and training strategies was developed to enhance the robustness of defect detection for small objects. The K-means++ clustering algorithm was used to update the initial values of the anchor frame to improve the performance of the model in detecting defects on the capsule surface. The experimental results showed that the proposed capsule defect detection method can accurately detect five types of defects which include dent, hole, scratch, stain and gap on its surface. Furthermore, the mean average precision of whether the capsule was defective or not was 99.05%, the average precision of each defect type was 91.81%, and the detection speed was 22 FPS. The method had certain advantages in comparison to other typical target detection methods in terms of detection speed and precision. The proposed YOLOv4-based defect detection method achieves the classification and localization of multiple types of defects in pharmaceutical hollow capsules, and has a better detection effect and stability, which can significantly reduce labor costs while fulfilling production quality control requirements.