基于深度学习的车型识别分析与研究

发布时间：2018-01-18 10:15

本文关键词：基于深度学习的车型识别分析与研究　出处：《山东师范大学》2017年硕士论文　论文类型：学位论文

【摘要】：随着城市交通复杂状况的愈渐加深,公安、交警的工作量也日益繁重,因而,智能交通系统的进一步发展变得尤为重要。车辆识别系统是目前比较热门且已经广泛应用于市场的一类智能交通系统,它依赖于车牌识别、车标识别和车型识别等核心技术。其中,车牌识别的研究已经比较成熟,许多研究成果在企业车库、小区等场所广泛应用。对于车标识别,由于车标区域太小,研究此类识别需要较高质量的图像,道路交通系统中的设备不能得到高质量的车辆图像的情况下,车标识别往往不能达到很好的识别效果。而传统的车型识别只能将车辆分成货车、客车、小汽车等几个大类,随着智能化发展愈来愈快,这类系统已经不能满足交通系统的需要,对如何识别车辆具体型号的研究在实际应用中将会有重要价值。类比于人脸识别,本文围绕车脸图像的识别进行了一系列研究。首先,由于缺少直接用于研究的车脸图像数据,本文截取车辆图像中的车脸部分进行尺度归一化,使用平移旋转、亮度调整、运动模糊等操作模拟现实中可能出现的几种因素进行数据扩增,保证数据的多样性,最终建立了一个包含31个子型号的车脸图像数据库。基于图像处理模式的图像识别方式,其核心是检测提取图像特征的算法,通过分析比较各种常用的特征提取算法,本文选择了在车辆识别领域最常用也是效果最好的SIFT特征提取算法,详细阐述了它的工作原理,并针对其生成的描述子过于复杂的缺点,在检测极值点时引入SUSAN角点检测算子进行改进,成功提取到了较简单且准确的车脸特征并结合k-邻近分类器进行分类。最后,本文引入了目前深度学习领域非常热门的卷积神经网络对车脸图像的特征进行自主学习、提取。首先是构建卷积神经网络模型,本文从网络层数、卷积核大小、下采样方法、激活函数类型等方面入手,分析不同的网络设定情况下卷积神经网络的训练时间、特征维度、特征提取时间等系统评价参数的变化,从中选取最优值,确定出卷积神经网络模型,将本文图像数据库中的图像输入网络,提取到最终的特征向量,输入k-邻近分类器进行实验,最终得出了较为理想的识别结果。与SIFT算法、SUSAN-SIFT算法相比,本文基于卷积神经网络的车型识别方法在速度与准确率方面都有所提升。
[Abstract]:With the city traffic situation more complex gradually deepened, public security, traffic police work is also increasingly heavy, therefore, the further development of the intelligent transportation system has become particularly important. Vehicle recognition system is currently more popular and has been widely used in the market for a class of intelligent transportation system, it depends on the license plate recognition, vehicle logo recognition and vehicle recognition core study on license plate recognition technology. Among them, have been mature, many research achievements in enterprises widely used in residential garage, etc. for vehicle logo recognition, the logo area is too small, the image of this recognition requires a higher quality of the road traffic system of the equipment can be vehicle images of high quality under the condition of vehicle logo recognition often can not achieve good recognition effect. And the traditional vehicle recognition only the vehicle is divided into passenger cars, trucks, and several other categories, with the development of intelligent More and more, this kind of system has been unable to meet the needs of the transportation system, have important value to study how to identify the vehicle specific models in practical applications. In analogy to the face recognition, the face image recognition around the car carried out a series of research. First of all, because of the lack of direct data on the face image for the car, car face this part of the interception of the vehicle image are normalized using scale, translation and rotation, brightness adjustment, motion blur operation simulation, several factors may occur in the reality of data amplification, ensure the diversity of the data, finally established a consists of 31 sub models face image database. The image recognition method based on image processing mode. Its core is the detection of image feature extraction algorithm, through the analysis and comparison of various commonly used feature extraction algorithm, we choose the most commonly used also in vehicle identification SIFT is the best feature extraction algorithm, expounds its working principle, and the generation of descriptors complicated, SUSAN corner detection operator is introduced in the detection limit point is improved and the successful extraction of the car face feature is simple and accurate and combined classifier. Finally, near k- at present, this paper introduces the characteristics of deep learning convolutional neural network is very popular field of car face image for autonomous learning extraction. First to construct the convolutional neural network model, this article from the network layer, the convolution kernel size, sampling method, activation function types and other aspects, analysis of different network settings for convolutional neural network case the training time, dimensionality, change feature extraction time and system parameters, selects the optimal value in determining the convolution neural network model, this paper image database The image input of the network, to extract the final feature vector, input k- neighbor classifier experiment, finally obtained satisfactory recognition results. Compared with the SIFT algorithm, SUSAN-SIFT algorithm, the vehicle recognition method based on convolutional neural network in speed and accuracy are improved.

【学位授予单位】：山东师范大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：U495;TP391.41

【参考文献】