基于卷积神经网络车型分类的研究

Research on vehicle classification based on Convolutional Neural Network

自ILSVRC大赛以来,卷积神经网络(Convolutional Neural Network,CNN)得到了迅速发展,众多学者将该技术用于图像的分类领域。车型分类是图像分类任务之一,在交通安全中具有很大的作用,因此使用CNN构建高效车型分类模型也越来越重要。为快速训练出高效的车型分类模型,首先使用迁移学习来训练本文的原始车型数据,其中InceptionV3模型精度最高,约85.91%。然而这些模型结构网络层次大多太深,且无法直接用于工程实践。因此,为了训练出精简且高效的模型结构,本文从CNN的基本概念出发,构建由卷积层、批规范层(Batch Normalization,BN)、池化层、Dropout层、全连接层和softmax层所组成的模型结构,使用带有约束权重的L2作为损失函数,并通过Adam优化算法对模型参数进行更新,又通过逐步增加卷积层数和调整全连接层神经元个数的方法,对数据增强过的车型数据进行训练和测试,结果表明卷积层数为4和全连接层神经元个数为256的模型结构的精度最好,约85.15%,较浅层次的网络达到了深层网络的性能。

Since the ILSVRC competition,the Convolutional Neural Network(CNN)has developed rapidly,and many scholars have used this technology in the field of image classification.Vehicle classification is one of the image classification tasks and plays an important role in traffic safety.Therefore,it is more and more important to use CNN to build an efficient vehicle classification model.In order to train an efficient vehicle classification model quickly,the paper uses the transfer learning to train the original data of this vehicle at first.As a result,the InceptionV3 of transfer learning model has the highest accuracy.It’s about 85.91%.However,the network structure of these models are too deep to be directly used in engineering practice.Therefore,in order to train a simple and efficient model structure,the paper starts the study based on the basic concept of CNN and builds the model by combinating the convolution layer,the batch normalization(BN),the pooling layer,the dropout layer,the fully connected layer and the softmax layer.Then the paper also uses the constraint weight with L2 as the loss function and update the model parameters by using the Adam optimization algorithm.Finally the paper trains the model by gradually increasing the number of convolution layers and adjusting the number of neurons in the fully connected layer with the data which is augmentated.The results show that the model structure in which the number of convolution layers is four and the fully connected layers’neurons are 256 is the best.It’s about 85.15%.The ability is almost equal to the ability of the deeper networks.