目的 为了准确而快速地自动检测医药玻璃瓶的外观缺陷。方法 基于YOLOv3算法，建立深度卷积神经网络学习检测模型，通过将神经网络结构浅层和深层特征向量连接并标准化，提取玻璃瓶多尺度特征信息。采用K-means聚类方式获得锚点框初始大小，提高模型对边界框预测的准确性；利用多尺度训练策略，增强模型对不同尺寸图像的鲁棒性。结果 实验结果表明，提出的医药玻璃瓶缺陷检测方法能够准确检测识别玻璃瓶上的管端残损、气线、气泡、划伤、污渍和结石等缺陷种类。与主流的目标检测方法相比，提出的方法在处理速度和准确度上都有提高，缺陷目标检测精确率达到96.23%，召回率为93.82%，平均精度为89.35%。结论 该方法已经在国内几家大型医药玻璃包装生产公司成功应用，显著提高了医药玻璃包装产品的质量和合格率，降低了人工成本。

The work aims to automatically detect the defects on the medicinal glass bottles accurately and quickly. Based on the YOLOv3 algorithm, a learning detection model of deep convolutional neural network was built. The multi-scale feature information of glass bottles was extracted by connecting and normalizing the shallow and deep feature vectors of the neural network structure. To improve the accuracy of the model for the prediction of the bounding box, K-means clustering method was used to obtain the initial size of the anchor box. The multi-scale training strategy was used to enhance the robustness of the model to images of different sizes. The experimental results showed that, the proposed defect detection method of medical glass bottles could accurately detect and identify such defects as damaged tube end, gas lines, bubbles, scratches, stains and calculi on glass bottles. Compared with the mainstream target detection methods, the processing speed and accuracy of the proposed method were improved. The accuracy of defect target detection reached 96.23%, the recall rate was 93.82%, and the average accuracy was 89.35%. The proposed method has been successfully applied to several large medical glass packaging production companies in China, which has significantly improved the quality and qualified rate of medical glass packaging products and reduced the labor costs.