摘要
针对直接集成简单分类器对交通标志数据库进行识别出现的类别预测效果较差的问题,提出一种基于卷积神经网络(CNN)和Bagging集成学习的交通标志识别算法,采用爬虫和图像增强技术实现交通标志数据集的扩充,以CNN网络提取交通标志图像的特征,通过采用最大池化层实现图像数据下采样,采用较浅的网络深度以简化整体网络结构。在CNN网络特征提取的基础上,利用软投票机制对多项Logistic、K近邻、SVM个体学习器进行集成,实现较准确的交通标志识别。实验结果表明,该算法在TSRD交通标志识别数据库测试集上的识别准确率达到了93.00%,相对于未改进的卷积神经网络模型识别准确率提高了11.99个百分点,并较高于通过VGG16和ResNet50迁移学习实现的识别准确率,具有较快的收敛速度。
To solve the problem of poor category prediction effect of direct integration of simple classifiers to recognize traffic sign databases,a traffic sign recognition algorithm based on Convolutional Neural Network and Bagging ensemble learning is proposed,using crawler and image enhancement technology to expand the data set of traffic signs,and the features of traffic sign images is extracted by Convolutional Neural Network.For the characteristics of the logo image,the image data are down-sampled by using the maximum pooling layer,and the overall network structure is simplified by using a shallower network depth.On the basis of Convolutional Neural Network feature extraction,the soft voting mechanism is used to integrate Multiple Logistic,K nearest Neighbor,and SVM individual learners to obtain a better prediction result and achieve more accurate traffic sign recognition.Experimental results show that the recognition accuracy of the algorithm on the TSRD traffic sign recognition database test set has reached 93.00%,which is 11.99%higher than the recognition accuracy of the original Network model,and is higher than the recognition accuracy achieved by VGG16 and ResNet50 migration learning having a faster convergence speed.
作者
田晟
宋霖
TIAN Sheng;SONG Lin(School of Civil Engineering and Transportation,South China University of Technology,Guangzhou Guangdong 510641,China)
出处
《广西师范大学学报（自然科学版）》
CAS
北大核心
2022年第4期35-46,共12页
Journal of Guangxi Normal University:Natural Science Edition
基金
广东自然科学基金(2021A1515011587,2020A1515010382)。
关键词
图像增强
卷积神经网络
迁移学习
集成学习
交通标志识别
image enhancement
convolutional neural network
transfer learning
integrated learning
traffic sign recognition
