简川霞,林子嘉,杜美剑,吴一凡,谢俊生.基于特征融合与降维的印刷套准识别方法[J].包装工程,2019,40(21):242-249.
JIAN Chuan-xia,LIN Zi-jia,DU Mei-jian,WU Yi-fan,XIE Jun-sheng.Printing Registration Recognition Method Based on Feature Fusion and Dimension Reduction[J].Packaging Engineering,2019,40(21):242-249. |
| 基于特征融合与降维的印刷套准识别方法 |
| Printing Registration Recognition Method Based on Feature Fusion and Dimension Reduction |
| 投稿时间：2019-07-05 修订日期：2019-11-10 |
| DOI：10.19554/j.cnki.1001-3563.2019.21.036 |
| 中文关键词: 印刷套准 特征降维 支持向量机 |
| 英文关键词: printing registration feature dimension reduction support vector machine |
| 基金项目:广东工业大学青年基金重点项目（17QNZD001）；广东省信息物理融合系统重点实验室项目（2016B030301008）；广东省数控一代机械产品创新应用示范工程专项资金项目（2013B011301023）；大学生创新创业训练计划（201911845008，xj201911845005，xj201911845014） |
|
| 摘要点击次数: |
| 全文下载次数: |
| 中文摘要: |
| 目的 针对单一方面特征难以准确表达印刷标志套准状态的问题，研究印刷标志图像多维特征提取、融合和降维的印刷套准识别方法。方法 提取印刷标志图像的灰度共生矩阵、Tamura纹理特征、灰度差分统计特征和灰度梯度共生矩阵表达其纹理，并采用主成分分析法对融合后的多维特征进行降维处理，得到主特征。将印刷标志图像的主特征数据分成训练集和测试集。支持向量机模型通过对训练集的学习确定模型参数，然后在测试集上验证模型的性能。结果 文中建议方法在测试集上的识别准确率为99%，训练集对支持向量机模型的训练时间为1.9327 s，模型在测试集上的识别时间为0.0307 s，模型的总体时间（训练时间和识别时间之和）为1.9634 s。结论 文中建议方法优于采用单一方面特征的识别准确率；同时在不影响识别准确率的情况下，优于未PCA降维方法的模型训练时间、识别时间和总体时间。 |
| 英文摘要: |
| The work aims to study the method of printing registration recognition which consists of the multi-dimension feature extraction, fusion and dimension reduction of the printing mark images, with respect to the problem of being unable to accurately represent the printing mark registration state with the single-style features. The gray level co-occurrence matrix, the Tamura texture feature, the gray difference statistical feature and the gray gradient co-occurrence matrix of the printing mark images were extracted to represent their texture. Then, the principal component analysis was carried out to reduce the dimension of the fused multi-dimensional features to obtain the principal features. The printing mark images with the principal featrues were divided into two sections: the training set and the testing set. The training set was learned by the support vector machine (SVM) model, so as to determine the parameters of this model, and the performance of this model was verified on the testing set. The proposed method achieved the recognition accuracy of 99% on the testing set, the SVM model training time of 1.9327 s on the training set, the SVM model recognition time of 0.0307 s on the testing set, and the model's total time (sum of training time and recognition time) of 1.9634 s. The proposed method outperforms the methods of the single-style features in terms of the recognition accuracy. Meanwhile, without decreasing the recognition accuracy, the proposed method is also better than the Non-PCA dimension reduction method in terms of the training time, the testing time and the total time. |
|
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