计算机工程与应用 ›› 2018, Vol. 54 ›› Issue (10): 19-25.DOI: 10.3778/j.issn.1002-8331.1801-0269

• 热点与综述 • 上一篇    下一篇

基于改进深度孪生网络的分类器及其应用

沈  雁1,王  环1,2,戴瑜兴1,2   

  1. 1.湖南大学 电气与信息工程学院,长沙 410082
    2.温州大学 数理与电子信息工程学院,浙江 温州 325035
  • 出版日期:2018-05-15 发布日期:2018-05-28

Deep siamese network-based classifier and its application

SHEN Yan1, WANG Huan1,2, DAI Yuxing1,2   

  1. 1.College of Electrical and Information Engineering, Hunan University, Changsha 410082, China
    2.College of Mathematics, Physics and Electronic Information Engineering, Wenzhou University, Wenzhou, Zhejiang 325035, China
  • Online:2018-05-15 Published:2018-05-28

摘要: 孪生神经网络由两组共享参数的孪生神经网络组成,可对高维度非线性的数据进行低维度映射,其在低维特征空间中变得可分。利用其优异的相似度计算性能,针对像交通标志识别这样具有复杂环境条件的分类问题,提出并设计基于孪生神经网络结构的高效分类器。采用卷积神经网络作为其基本构成,运用max-pooling,dropout等技术形成特征提取所需的多尺度卷积神经网络。同时辅助以空间变换器网络来进一步提高识别的准确率。通过对GTSRB交通标志数据集进行测试,其识别的准确率达到了99.40%。该分类器方法同时具备了结构简单、训练时间短、准确率高以及识别速度快的优点。

关键词: 孪生神经网络, 分类器, 空间变换器网络, 交通标志识别

Abstract: Siamese neural network consists of twin networks which share the same parameters, and can map nonlinear data with high dimension onto a low dimension and easy separable feature space. By taking use of its excellent performance on similarity computing, an efficient classifier based on Siamese network is proposed. Convolutional Neural Networks(CNNs) are used as building blocks, combined with max-pooling and dropout techniques, to construct a multi-scale CNNs for feature extraction. A spatial transformer network is used as a supplementary to promote the accuracy of the classifier. Through experiments on traffic sign data GTSRB, the classifier obtains 99.40% accuracy rate for this recognition benchmark. The proposed method for classification has virtues of simple structure, high accuracy, short training and fast recognition.

Key words: Siamese network, classifier, spatial transformer networks, traffic sign recognition