简川霞,高健.面向不均衡训练集的印刷图像套准状态检测方法[J].包装工程,2018,39(11):158-164.
JIAN Chuan-xia,GAO Jian.Registration Recognition Methods of the Printing Images Oriented to the Imbalanced Training Set[J].Packaging Engineering,2018,39(11):158-164. |
| 面向不均衡训练集的印刷图像套准状态检测方法 |
| Registration Recognition Methods of the Printing Images Oriented to the Imbalanced Training Set |
| 投稿时间：2018-01-28 修订日期：2018-06-10 |
| DOI：10.19554/j.cnki.1001-3563.2018.11.028 |
| 中文关键词: 不均衡数据 印刷套准 集成采样 支持向量机 |
| 英文关键词: imbalanced data printing registration integrated sampling support vector machine |
| 基金项目:国家自然科学基金（51675106）；广东省自然科学基金（2015A030312008, 2016A030308016）；广东省科技计划（2015B010104008）；广东工业大学青年基金重点项目（17QNZD001） |
|
| 摘要点击次数: |
| 全文下载次数: |
| 中文摘要: |
| 目的 针对不均衡的印刷图像套准状态检测中存在的印刷套不准图像识别准确率低的问题，研究不均衡印刷图像训练集的预处理方法。方法 提出不均衡印刷图像训练集数据的集成采样预处理方法。支持向量机先将不均衡的训练集数据分为支持向量和非支持向量，然后过采集少类样本（即印刷套不准图像）中的支持向量，欠采集多类样本（即印刷套准图像）中的非支持向量，实现训练集数据的均衡化。最后采用预处理后的均衡训练集对支持向量机模型进行训练，并优化模型参数。结果 采用文中提出的集成采样方法对不均衡训练集预处理后获得支持向量机模型，通过对印刷图像套准状态进行识别，获得的少类样本识别率a+为0.9375，识别准确率几何平均数Gmean为0.9437，F测度为0.9574。结论 文中提出方法获得的印刷套不准图像识别准确率a+, Gmean和F测度均优于实验中的其他方法。 |
| 英文摘要: |
| The work aims to study a preprocessing method for the imbalanced printing image training set in view of the problem of poor recognition accuracy of unaligned printing images in the detection of imbalanced printing image registration. An integrated sampling preprocessing method for the imbalanced printing image training set was proposed. Firstly, the imbalanced training set was divided into support vectors and non-support vectors through the support vector machine model. Secondly, in order to balance the training set, the support vectors in the minority class (i.e., the unaligned printing images) were oversampled and the non-support vectors in the majority class (i.e., the aligned printing images) were undersampled. Finally, the pre-processed balanced training set was used to train the support vector machine model, and the model parameters were optimized. The proposed integrated sampling method was used to pre-process the imbalanced training set to obtain the support vector machine model. The recognition rate of the minority class a+ obtained through the recognition of printing image registration was 0.9375, the geometric mean Gmean of the recognition accuracy was 0.9437 and the F-score was 0.9574. The proposed method outperforms other methods in the experiment in terms of the recognition accuracy a+ of the unaligned printing images, Gmean and F-score. |
|
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