摘要
喷嘴堵塞是数码喷印过程中常见的问题,为了能够及时发现、避免大量产品报废,需要对生产过程中喷印图像进行实时检测。针对目前数码喷印缺陷检测传统方法效率低,检测速度与检测精度不平衡的情况,本研究提出了一种轻量级语义分割的缺陷检测方法。使用GhostNet对Deeplabv3+中的主干网络进行轻量化网络结构改进,在Deeplabv3+的ASPP模块中加入3个残差模块,对空洞卷积进行上下文结构调整,并使用Focal Loss损失函数和ReLU6激活函数。实验结果表明,改进后的模型Gh-R-Deeplabv3+在数码喷印缺陷数据集上每秒处理帧数为47.71,并达到了82.8%的平均交并比以及90.96%的平均像素准确率,取得了较高的检测速度和检测精度,并符合实时检测标准。证实了本研究模型对数码喷印缺陷在检测速度和检测精度上的有效性。
Ink jet nozzle clogging is a common problem that occurs in the process of digital ink-jet printing.Real-time detection of ink-jet printing images is necessary to identify and promptly avoid the scrapping of a large number of products during the production process.Given the inefficiency of the current traditional methods for defect detection in digital ink jet printing and the imbalance between detection speed and accuracy,a lightweight semantic segmentation defect detection method had been proposed.The backbone network structure of Deeplabv3+was improved using GhostNet,three residual modules were added to the ASPP module in Deeplabv3+,the context structure of the cavity convolution was adjusted,and the Focal Loss loss function and ReLU6 activation function were utilized.The experimental results demonstrated that the improved model,Gh-R-Deeplabv3+,processed 47.71 frames per second on the digital ink jet printing defect data set.It achieved an average cross merge ratio of 82.8%and an average pixel accuracy of 90.96%.The model was confirmed to have achieved a high detection speed and accuracy,conforming to the real-time detection standard.It was confirmed that a relative balance between detection speed and accuracy for digital ink jet printing defects is achieved by the improved model in this study.
作者
肖蕾
张铭芷
李琪
陈镇家
XIAO Lei;ZHANG Ming-zhi;LI Qi;CHEN Zhen-jia(College of Automation,Guangdong Polytechnic Normal University,Guangzhou 510665,China)
出处
《印刷与数字媒体技术研究》
CAS
北大核心
2024年第4期218-227,共10页
Printing and Digital Media Technology Study
