卷烟包装外观缺陷数据集构建及深度学习检测技术研究

目的 为了提升烟包缺陷检测的准确率，构建卷烟包装外观缺陷识别基准数据集，并开展主流深度学习模型在卷烟包装外观缺陷智能检测中的应用研究。方法 首先，从生产运行中的ZB45型细支烟硬盒包装机组采集缺陷图像，经过人工审核与筛选后获取典型的缺陷数据。然后，根据缺陷的特征与成因，将缺陷数据划分为23个类别，并逐一进行目标检测框标注。最终，形成了包含13 000余张缺陷图像的卷烟包装外观缺陷识别基准数据集，并针对烟包缺陷识别、缺陷分类、目标检测、模型迁移4项任务开展实验。结果 结果表明，数据集能够满足高准确率深度学习模型的训练需求；通过模型迁移，能够利用该数据集大幅提高不同牌号卷烟的缺陷检测效果；DenseNet模型在烟包缺陷识别与缺陷分类任务上表现较好，准确率分别达到93.70%和95.43%,YOLOv5模型在缺陷目标检测任务上mAP@0.5值达到了96.61%。结论 该数据集能够作为烟包缺陷检测领域的基准数据集，研究成果将进一步支撑卷烟包装领域的数据应用与数字化转型。

Cigarette Packaging Appearance Defect Data Set Construction and Deep Learning Detection Technology Research

The work aims to construct a benchmark dataset for cigarette package appearance defect recognition and carry out research on the application of mainstream deep learning models in the intelligent detection of cigarette package appearance defects, so as to improve the accuracy of cigarette package defect detection. The image data of suspected defects were collected from the normal production ZB45 fine cigarettes hard box packaging machine, and the data with respect to real defects were obtained through manual reviews and screening. According to the characteristics and causes of defects, the defect data were classified into 23 categories, the labels and locations of defect were marked with bounding boxes. A benchmark dataset containing more than 13 000 images of cigarette package appearance quality defects was constructed. Experimental tests were conducted for four tasks, namely, cigarette package defect recognition, defect classification, target detection, and model transfer. The results showed that the dataset fulfilled the training requisites for high-accuracy deep learning models; Through model migration, the dataset could be utilized to significantly improve the accuracy of defect detection for different cigarette grades; The DenseNet model achieved better results on the cigarette packet defect recognition and defect classification tasks, with accuracy rates of 93.70% and 95.43%, respectively, and the YOLOv5 model achieved a mAP@0.5 of 96.61% on the defective target detection task. The dataset can be used as a benchmark dataset in cigarette packet defect detection, and the research results will further support the data application and digital transformation in cigarette packaging.