基于卷积神经网络的道路监控系统下车辆颜色识别  

Vehicle Color Recognition in Road Monitoring System Based on Convolution Neural Network

作　　者：姚国愉 张昭 李雪纯 张佳乐 YAO Guoyu;ZHANG Zhao;LI Xuechun;ZHANG Jiale

机构地区：[1]西安邮电大学通信与信息工程学院,陕西西安710121

出　　处：《科技创新与应用》2021年第8期86-89,共4页Technology Innovation and Application

摘　　要：针对卷积神经网络训练时收敛速度慢且参数数量较多的问题,文章在激活函数之前使用批归一化对每一个小批量数据进行处理,并使用1x1的卷积层和全局平均池化层代替全连接层,提出了一种基于卷积神经网络的车辆颜色识别方法。该方法是专门为识别任务而设计的,它包含八层,分别是五个卷积层,两个1x1的卷积层和一个全局平均池化层。实验结果表明,文章在训练集上的识别精度为99.6%,在测试集上的识别精度为94.8%,与现存最优的实验结果相比,识别精度提高了0.33%,且参数量仅占其14.5%。Aiming at the number of parameters and slow convergence of training in convolution neural network,in this paper,Batch Normalization operation for each small batch data is used before activating function,1×1 convolutional layer and a global average pooling layer is employed to instead of a fully connected layer,and then a vehicle color recognition method using convolutional neural network is proposed.This method is specially designed for recognition tasks,which contains eight layers,namely five convolutional layers,two convolutional layers of 1×1 and a global average pooling layer.The experimental results show that,compared with the state-of-art method,the proposed method can obtain more than 0.33%higher recognition accuracy on the test set,up to 94.8%,and can also obtain 99.6%recognition accuracy on the training set,in which the number of parameters is only 14.5%.

关 键 词：批归一化 卷积神经网络 车辆颜色识别 识别精度 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]