张志晟,张雷洪.基于深度学习的易拉罐缺陷检测技术[J].包装工程,2020,41(19):259-266.
ZHANG Zhi-sheng,ZHANG Lei-hong.Defect Detection Technology for Cans Based on Deep Learning[J].Packaging Engineering,2020,41(19):259-266. |
| 基于深度学习的易拉罐缺陷检测技术 |
| Defect Detection Technology for Cans Based on Deep Learning |
| 投稿时间：2020-03-04 修订日期：2020-10-10 |
| DOI：10.19554/j.cnki.1001-3563.2020.19.037 |
| 中文关键词: 缺陷检测 图像评估 深度学习 迁移学习 图像处理 |
| 英文关键词: defect detection image evaluation deep learning transfer learning image processing |
| 基金项目: |
|
| 摘要点击次数: |
| 全文下载次数: |
| 中文摘要: |
| 目的 现有的易拉罐缺陷检测系统在高速生产线中存在错检率和漏检率高，检测精度相对较低等问题，为了提高易拉罐缺陷识别的准确性，使易拉罐生产线实现进一步自动化、智能化，基于深度学习技术和迁移学习技术，提出一种适用于易拉罐制造的在线检测的算法。方法 利用深度卷积网络提取易拉罐缺陷特征，通过优化卷积核，减短易拉罐缺陷检测的时间。针对国内外数据集缺乏食品包装制造的缺陷图像，构建易拉罐缺陷数据集，结合预训练网络，通过调整VGG16提升对易拉罐缺陷的识别准确率。结果 对易拉罐数据集在卷积神经网络、迁移学习和调整后的预训练网络进行了易拉罐缺陷检测的性能对比，验证了基于深度学习的易拉罐缺陷检测技术在学习率为0.0005，训练10个迭代后可达到较好的识别效果，最终二分类缺陷识别率为99.7%，算法耗时119 ms。结论 相较于现有的易拉罐检测算法，文中提出的基于深度学习的易拉罐检测算法的识别性能更优，智能化程度更高。同时，该研究有助于制罐企业利用深度学习等AI技术促进智能化生产，减少人力成本，符合国家制造业产业升级的策略，具有一定的实际意义。 |
| 英文摘要: |
| In allusion to the problem that the existing can defect detection system has high false detection rate and missed detection rate and relatively low detection accuracy in high-speed production lines, this paper aims to propose an algorithm for online detection of can manufacturing based on deep learning and transfer learning techniques to improve the accuracy of can defect identification and make the can production line more automated and intelligent. This paper used deep convolutional network to extract the characteristics of can defects, and optimized the convolution kernel to reduce the time required for can defect detection. Due to the lack of domestic and foreign data sets for defective images of food packaging manufacturing, a can defect data set was built, and the accuracy of identifying can defects was improved by adjusting VGG16 in combination with pre-trained network. The performance comparison for can test was performed on the convolution neural network, transfer learning, and adjusted pre-trained network. It was verified that this technology had a good recognition effect when the learning rate was 0.0005 and the epochs of trainings was 10. The recognition rate of the final binary classification was 99.7% and the time consumption of algorithm was 119 ms. Compared with the existing can detection algorithms, this paper proposes a deep learning-based can detection algorithm with better recognition performance and higher intelligence. At the same time, this study helps enterprises to use AI technology such as deep learning to promote intelligent production, reduce human cost, and conform to the strategy of national manufacturing industry upgrading, which has certain practical significance. |
|
查看全文
查看/发表评论 下载PDF阅读器 |
| 关闭 |
渝公网安备 50010702501716号