欢迎访问《图学学报》 分享到:

图学学报 ›› 2021, Vol. 42 ›› Issue (1): 15-22.DOI: 10.11996/JG.j.2095-302X.2021010015

• 图像处理与计算机视觉 • 上一篇    下一篇

液晶字符识别的 CNN 和 SVM 组合分类器

  

  1. (沈阳理工大学信息科学与工程学院,辽宁 沈阳 110159)
  • 出版日期:2021-02-28 发布日期:2021-01-29
  • 基金资助:
    辽宁省自然科学基金项目(20170540792) 

A combined classifier based on CNN and SVM for LCD character recognition 

  1. (School of Information Science and Engineering, Shenyang Ligong University, Shenyang Liaoning 110159, China) 
  • Online:2021-02-28 Published:2021-01-29
  • Supported by:
    Natural Science Foundation of Liaoning Province (20170540792) 

摘要: 针对仪表液晶显示字符识别问题,提出一种结合了卷积神经网络(CNN)和支持向量机(SVM)的 字符识别方法。分别采用具有并联结构的 CNN 模型和基于梯度方向直方图(HOG)特征的 SVM 方法构建基本分 类器,当 2 个分类器的结果存在冲突时,利用 CNN 的 softmax 输出最大值判决最终结果,当其大于设定阈值 时采用 CNN 分类器的结果,反之采用 SVM 分类器的结果。建立字符图像的误差模型并利用仿真方法构建了 数据集用于分类器的训练和测试,给出一种基于投票原理的最优阈值的估计算法。在 MNIST 和仿真数据集上 的测试实验结果表明,最优阈值估计算法的结果可靠,组合分类器的准确率较 2 种单一分类器均有提高,在实 际测试系统上其准确率达到 99.81%,验证了该组合分类器方法对液晶字符识别问题的有效性;在 CIFAR-10 数 据集上的实验结果验证了该方法也可用于其他分类问题。

关键词: 计算机视觉, 机器学习, 液晶字符识别, 支持向量机, 卷积神经网络

Abstract: A combined classifier based on convolution neural network (CNN) and support vector machine (SVM) was proposed for the recognition of liquid crystal displayer (LCD) characters. Two basic classifiers were utilized to build a combined classifier for recognition. One was CNN with a parallel structure, and the other was SVM using histogram of oriented gradients (HOG) features of the character image. If a sample’s responses from two basic classifiers conflicted with each other, the maximum component of the softmax vector outputted from CNN classifier was employed to determine the final result. If it was greater than a threshold, the CNN result was adopted, otherwise the SVM result. An error model for LCD character image was presented and adopted to construct a simulation dataset for the algorithm training and test. An optimal threshold estimation algorithm based on voting principle was proposed. The combined classifier was tested on both MNIST dataset and an LCD character simulation dataset. The experimental results show that the threshold estimation result was reliable, and that the combined classifier outperformed both CNN and SVM basic classifiers. Using the method on a real test system, the accuracy rate was 99.81%. The results prove the method’s effectiveness for LCD character recognition. The experimental results on CIFAR-10 dataset show that the method can also be applied to other kinds of classifications. 

Key words: computer vision, machine learning, liquid crystal displayer character recognition, support vector machine, convolution neural network 

中图分类号: