朱立明,王伟,范霞萍,王文博,徐鑫,许小双.基于无监督深度神经网络的卷烟小包拉线缺陷视觉智能检测方法[J].包装工程,2022,43(17):273-281.
ZHU Li-ming,WANG Wei,FAN Xia-ping,WANG Wen-bo,XU Xin,XU Xiao-shuang.Vision Intelligent Detection Method of Cigarette Packet Tear Tape Defects Based on Unsupervised Deep Neural Network[J].Packaging Engineering,2022,43(17):273-281. |
| 基于无监督深度神经网络的卷烟小包拉线缺陷视觉智能检测方法 |
| Vision Intelligent Detection Method of Cigarette Packet Tear Tape Defects Based on Unsupervised Deep Neural Network |
| |
| DOI：10.19554/j.cnki.1001-3563.2022.17.036 |
| 中文关键词: 卷烟小包拉线 生成对抗网络 自编码器 视觉智能检测 |
| 英文关键词: cigarette packet tear tape generative adversarial networks autoencoder vision intelligent detection |
| 基金项目:中国烟草总公司科技项目(110202102006) |
|
| 摘要点击次数: |
| 全文下载次数: |
| 中文摘要: |
| 目的 为减少ZB47包装机小包拉线缺陷投诉，基于无监督深度神经网络构建一种小包拉线缺陷视觉智能检测方法。方法 首先，在ZB47包装机CH转塔部位设计并加装小包图像采集装置，获得实时高清晰度小包图像。其次，将小包图像根据拉线位置进行固定位置的裁剪，从而减轻不同工况的环境背景影响并且加快检测速度。然后，构建自编码器–编码器结构的主干网络，同时叠加生成对抗网络中的判别器模块组成缺陷判别模型，并综合采用图像间、图像隐空间以及图像特征间的信息构建模型的损失函数。最后，使用裁剪后的正常小包拉线图像对构建的缺陷判别模型进行训练，并基于所有的正常小包图像得到异常阈值。结果 实际验证阶段，待检测图像的得分大于异常阈值即判断为异常图像，触发CH转塔部位的小包剔除装置将该缺陷小包剔除。生产现场测试表明，所提方法可以对典型小包缺陷进行快速准确检测，缺陷检测准确率为99.99%。结论 该方法能够满足生产现场卷烟小包拉线缺陷检测的准确性和实时性要求。 |
| 英文摘要: |
| The work aims to construct a vision intelligent detection method for cigarette packet tear tape defects based on the unsupervised deep neural network to reduce the complaints of cigarette packet tear tape defects for the ZB47 packaging machine. First, the cigarette packet image acquisition hardware acquisition device at the CH turret position of the ZB47 packaging machine was designed and installed to obtain real-time high-precision small packet images. Second, the cigarette packet image was cropped at a fixed position according to the position of the tear tape, thereby reducing the effects of the environmental background of different working conditions and speeding up the detection speed. Then, the backbone network of the autoencoder-encoder structure was constructed, and the discriminator module in the generative adversarial networks was added to form the defect detection module. The loss function of the model was constructed according to the information between the images, the latent space and the features of the images. Finally, the cropped normal cigarette packet transparent paper images were used to train the constructed defect detection model, and the abnormal score threshold was obtained based on all normal cigarette packet images. In the actual verification stage, if the score of the detected image was greater than the abnormal score threshold, it is judged to be an abnormal image, and the cigarette packet removal device at the CH turret position was triggered to remove the defective cigarette packet. The test at production site showed that the proposed method could quickly and accurately detect the cigarette packet tear tape defects with an accuracy rate of 99.99%. The method can meet the dual requirements of the actual production process for detection accuracy and detection speed of cigarette packet tear tape defects. |
|
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