基于改进Faster-RCNN算法的软包装印刷缺陷检测研究  

Research on Printing Defect Detection of Flexible Packaging Based on Improved Faster-RCNN Algorithm

作　　者：马克西姆 郭蓉 Maxim;GUO Rong(Intelligent Manufacturing Laboratory,Beijing Institute of Graphic Communication,Beijing 102600,China)

机构地区：[1]北京印刷学院智能制造实验室,北京102600

出　　处：《北京印刷学院学报》2024年第3期28-31,44,共5页Journal of Beijing Institute of Graphic Communication

基　　金：北京市高教学会教改项目-数字孪生技术在自动化课程教学中的应用研究(22150223005)成果。

摘　　要：本文提出了一种改进的Faster-RCNN算法进行缺陷检测,针对软包装漏印的特点,将原来的VGG16网络替换成运算量更小、网络深度更深的残差网络(ResNet-50),可以提取丰富的特征。为了使卷积神经网络自适应注意,在ResNet-50的残差网络中添加了CBAM自注意力机制模块。对于数据集,对采集的图像通过旋转、平移、亮度调整、加入噪声、Cutout等操作进行数据增强,避免数据样本不均衡,提升模型的鲁棒性。结果显示,改进后的Faster-RCNN模型与未改进的Faster-RCNN模型相比准确率提高了12%,mAP达到92.95%。证明改进后模型的有效性,节省大量人工成本,提高企业生产效率。This paper proposes an improved Faster-RCNN algorithm for defect detection,and the original VGG16 network is replaced by a residual network(ResNet-50)with smaller computation and deeper network depth for the characteristics of missing prints in flexible packaging,which can extract rich features.In order to make the convolutional neural network adaptive attention,the CBAM selfattention mechanism module was added to the residual network of ResNet-50.For datasets,the collected images are enhanced by rotation,translation,brightness adjustment,noise addition,cutout,and other operations to avoid unbalanced data samples and improve the robustness of the model.The results show that the precision of improved Faster-RCNN model is 12%higher than the unimproved Faster-RCNN model,and the mAP reaches 92.95%.Prove the effectiveness of the improved model,save a lot of labor costs,and improve the production efficiency of the enterprise.

关 键 词：Faster-RCNN 漏印 目标检测 

分 类 号：TS807[轻工技术与工程]