基于CNN和SVM融合的交通标志识别

Traffic Sign Recognition Based on Combination of CNN and SVM

针对传统的卷积神经网络对小样本分类易产生过拟合等问题,在卷积神经网络(CNN)和支持向量机(SVM)融合模型的基础上,提出对CNN网络结构提取的特征进行归一化处理,提高泛化能力,并将其应用到交通标志识别。该方法构建了一种CNN-SVM模型,将卷积神经网络和支持向量机结合起来,使用从ImageNet数据集初始化的网络进行特定域的微调,截取网络内层来提取交通标志图像特征,并对特征进行归一化处理,最后采用SVM进行识别,从而有效解决交通标志分类过拟合问题。仿真结果表明,通过CNN内层建立的特征映射模型,所传递的特征经过归一化处理后,在交通标志分类任务中具有良好的特征表示能力,较好地提升了SVM分类性能,表现出更好的分类精度以及泛化性能。

Aiming at the problem that the traditional convolutional neural network tends to over-fit the classification of small samples,we propose a new method to normalize the features extracted from CNN network structure based on the model of CNN and SVM combination,so as to improve its generalization ability and apply it to traffic sign recognition.The method builds a CNN-SVM model which combines convolutional neural network and support vector machine,and fine-tunes the network initialized from ImageNet dataset on specific domain and intercepts its inner layer to extract the image features from traffic signs.The features are normalized and finally identified by SVM,which effectively solves the over-fitting problem toward traffic signs’classification.The simulation results show that the feature mapping model established through the inner layer of CNN,after the transmitted features being normalized,have a superb feature presentation ability in traffic sign classification tasks,improve the SVM classification performance,and show better classification accuracy and generalization performance.