计算机工程与应用 ›› 2021, Vol. 57 ›› Issue (11): 248-253.DOI: 10.3778/j.issn.1002-8331.2003-0111

• 工程与应用 • 上一篇    下一篇

融合独立组件的ResNet在细粒度车型识别中的应用

陈立潮,朝昕,曹建芳,潘理虎   

  1. 1.太原科技大学 计算机科学与技术学院,太原 030024
    2.忻州师范学院 计算机系,山西 忻州 034000
  • 出版日期:2021-06-01 发布日期:2021-05-31

Application of ResNet with Independent Components in Fine-Grained Vehicle Recognition

CHEN Lichao, CHAO Xin, CAO Jianfang, PAN Lihu   

  1. 1.School of Computer Science and Technology, Taiyuan University of Science and Technology, Taiyuan 030024, China
    2.Department of Computer, Xinzhou Teachers University, Xinzhou, Shanxi 034000, China
  • Online:2021-06-01 Published:2021-05-31

摘要:

针对细粒度车型中子车系间识别率低的问题,同时为了增强卷积神经网络的表征能力,提出融合独立组件的残差网络(IC-ResNet)模型。优化ResNet网络,通过改进下采样层,减少特征信息损失,接着使用中心损失函数和Softmax损失函数联合学习策略,增强模型的类内聚性。在卷积层前引入独立组件(IC)层,获得相对独立的神经元,增强网络独立性,提高模型的特征表示能力,从而对细粒度车型实现更准确的分类。仿真实验表明,该模型在Stanford cars-196数据集上的识别准确率达到94.7%,与其他模型相比,实现了最优效果,从而验证了该车型识别模型的有效性。

关键词: 细粒度车型识别, 残差网络, 独立组件, 中心损失

Abstract:

Aiming at the problem of low recognition rate between the same car series in fine-grained models, in order to enhance the representational ability of the convolutional neural network, a ResNet model with integrated Independent Components(IC-ResNet) is proposed. Firstly, ResNet is optimized to reduce the loss of feature information by improving the lower sampling layer, and then the center loss function and Softmax loss function are combined to improve the class cohesion of the model. Then an IC layer is introduced in front of the convolution layer to obtain relatively independent neurons, enhance network independence, and improve the feature representation ability of the model, so as to achieve more accurate classification of fine-grained vehicle models. The experiment shows that the model recognition accuracy on the Stanford cars-196 data set is 94.7%, which achieves the optimal effect compared with other models and verifies the effectiveness of the recognition model of this model.

Key words: fine-grained vehicle recognition, Residual Network(ResNet), Independent Component(IC), central loss