王胜,吕林涛,杨宏才.卷积神经网络在印刷品缺陷检测的应用[J].包装工程,2019,40(11):203-211.
WANG Sheng,LYU Lin-tao,YANG Hong-cai.Application of Convolutional Neural Network in Printing Defect Detection[J].Packaging Engineering,2019,40(11):203-211. |
| 卷积神经网络在印刷品缺陷检测的应用 |
| Application of Convolutional Neural Network in Printing Defect Detection |
| 投稿时间：2019-01-21 修订日期：2019-06-10 |
| DOI：10.19554/j.cnki.1001-3563.2019.11.031 |
| 中文关键词: 印刷品产品质量 印刷品缺陷检测 卷积神经网络 模式识别 |
| 英文关键词: quality of printing products defect detection of printing products convolutional neural network pattern recognition |
| 基金项目:国家自然科学基金（61273271）；2016年度陕西省工业科技攻关项目（2016GY-141）；2017年度西安市科技产学研项目（2017087CG/RC050(XJXY001)）；西京学院校级科研基金（XJ160232） |
|
| 摘要点击次数: |
| 全文下载次数: |
| 中文摘要: |
| 目的 为了改善传统机器检测印刷产品缺陷存在误费率高的不足。方法 提出以卷积神经网络为控制核心的印刷品缺陷检测系统。设计可在实际检测中应用的卷积神经网络，设计在线印刷质量检测系统的硬件结构。结果 对结构相同而训练次数、学习率不同的卷积神经网络进行了缺陷检测的性能对比，验证了该卷积神经网络在学习率小于0.01时，可以获得较好的识别效果；在学习率大于0.05时，网络不容易收敛。网络训练次数越多，精度越高，相应的训练时间也较长。在满足快速性和精确度的条件下，确定了适应某印刷品的缺陷检验网络训练次数为50，学习率为0.005，此时的识别率为90%。结论 经过实验证明，该检测系统具有良好的缺陷识别能力，缺陷类型的分类准确率较高。该系统具有一定的实用价值。 |
| 英文摘要: |
| The paper aims to overcome the shortcoming of high error rate in traditional machine inspection of printing defects. A printing defect detection system based on convolutional neural network was proposed. The convolution neural network which can be used in practical inspection was designed, and the hardware structure of online printing quality detection system was designed. The defect detection performance of convolutional neural network with the same structure but different training times and different learning rate parameters was compared. It was verified that convolutional neural network with learning rate below 0.01 can achieve better recognition effect, while the learning rate network is hard to be converged when the learning rate is above 0.05. The more the number of network training, the higher the accuracy, and the longer the training time correspondingly. Under the premise of meeting the requirements on rapidity and accuracy, the number of network training for defect inspection of a printed matter was determined to be 50 times, and the learning rate was 0.005. In this case, the recognition rate was 90%. Experiments show that the detection system has good defect recognition ability and high classification accuracy of defect types. The system has certain practical value. |
|
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