卷积神经网络在交通标志识别中的研究

发布时间：2018-04-26 20:57

本文选题：交通标志识别 + Siamese网络　；参考：《广西师范大学》2017年硕士论文

【摘要】：自然环境下的交通标志识别,是无人驾驶中必不可少的技术要求,近年来百度、亚马逊、google、facebook等公司越来越重视无人驾驶,前几年的研究者通常应用颜色、形状、尺度等来表达交通标志的特征。本文充分利用最近5年热起来的卷积神经网络来识别交通标志,把网络不断优化,多次实验,调整网络结构与参数,最终设计了一个在交通标志识别中表现优秀的网络模型,该网络具有良好的准确性、高效性、可移植性。用卷积神经网络的理论解决分类识别问题时,往往用到大量的数据,本文是基于德国交通标志数据集(GTSRB),以在图像处理领域效果最好的Caffe作为框架,设计适合于交通标志识别的网络模型以及网络中使用的超参数。本文的研究工作主要包含以下几个方面:(1)本文在Siamese网络模型的基础上,提出了一种双网络线性合并的想法,首先将Siamese网络卷积层与池化层的权值共享的特性改进成先不共享,随着图片特征的不断提取,再进行权值共享,这样可以提高交通标志的准确率。(2)为了进一步优化网络的结构,提高交通标志识别的准确性与可移植性,本文采用2-channel网络结构思想,把单通道图片改进为双通道图片,该方法提高了网络模型的泛化能力,还能提取图像的特征表达能力,并且给交通标志识别中因为数据量的不足,带来更加方便的识别方法。(3)最后本文应用了全新的Inception结构为基本元素,以Ubuntu14.04为操作系统,以Caffe框架为平台,以GoogleNet.为整体网络模型构架,设计了一个高精度的TrafficNet网络模型,主要改进了网络的层级结构与激励函数,大大加快了网络训练的速率,针对43类交通标志,总计54324张图片最终取得了 99.8%的分类正确率。
[Abstract]:Traffic sign recognition in the natural environment is an essential technical requirement in driverless driving. In recent years, companies such as Baidu, Amazon Google Facebook and other companies have increasingly attached importance to driverless driving. In previous years, researchers often used colors and shapes. Scale, etc., to express the characteristics of traffic signs. In this paper, we make full use of convolutional neural network, which is hot in the last five years, to identify traffic signs, optimize the network, experiment many times, adjust the network structure and parameters, and finally design a network model with excellent performance in traffic sign recognition. The network has good accuracy, high efficiency and portability. In order to solve the problem of classification and recognition by convolution neural network theory, a lot of data are often used. This paper is based on the German traffic sign data set GTSRB, and takes Caffe, which has the best effect in the field of image processing, as the framework. The network model suitable for traffic sign recognition and the super parameters used in the network are designed. The research work of this paper mainly includes the following several aspects: 1) based on the Siamese network model, this paper proposes an idea of linear combination of double networks. Firstly, the characteristics of weight sharing between the convolutional layer and the pool layer of Siamese network are improved to not be shared first. With the continuous extraction of image features and the sharing of weights, the accuracy of traffic signs can be improved. In order to further optimize the network structure and improve the accuracy and portability of traffic sign recognition, In this paper, the idea of 2-channel network structure is adopted to improve the single channel image into two channel images. This method improves the generalization ability of the network model, and can extract the feature expression ability of the image, and it can also give the traffic sign recognition because of the shortage of data quantity. Finally, this paper uses a new Inception structure as the basic element, Ubuntu14.04 as the operating system, Caffe framework as the platform, and Google Net. A high-precision TrafficNet network model is designed for the overall network model architecture, which mainly improves the hierarchical structure and incentive function of the network, greatly speeds up the training rate of the network, aiming at 43 kinds of traffic signs. A total of 54324 images finally achieved 99.8% of the classification accuracy.
【学位授予单位】：广西师范大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41;TP183

【相似文献】