【作者】 方杰；

【导师】 都思丹；

【作者基本信息】 南京大学 ， 电路与系统， 2018， 硕士

【摘要】 车型细类识别是智能交通系统领域中的一项重要研究课题,同时也属于计算机视觉中的一种任务。车型细类识别即通过从车辆图片或者视频中识别出车辆的具体型号。不同于传统的车型识别,即识别车辆的种类,例如轿车,卡车,客车等等,车型细类识别将识别出具体生产厂商的某一种具体车型。传统的车型细类识别研究主要通过人工设计特征和基于3D模型等方法来进行,准确率和泛化性较差,较难应用到实际的智能交通系统中。近年来,深度学习尤其是卷积神经网络在人工智能领域取得了巨大的成功,在图像识别领域,目前是占据统治地位的方法。它不仅仅极大地提高了识别精确度,同时减少了人工设计特征的复杂性。现在,一些车型细类识别的任务也逐渐开始使用卷积神经网络。本文提出了一个基于卷积神经网络的从粗糙到精细的车型细类识别系统。在这个系统中,我们将利用从大规模数据集中提取的车型的全局特征和最精细的局部特征来进行车型识别。为了得到最精细的局部特征,我们提出了两个重要算法,一个是定位局部区域算法,一个是从粗糙到精细的检测算法。对于定位局部区域算法,我们将建立一个从卷积神经网络得到特征热点图到输入数据之间的映射关系,从而根据热点图中热点区域提取出输入数据中最具有判别信息的部分。对于从粗糙到精细的检测算法,我们将反复利用定位局部区域算法对提取出来的部分再次进行更精细部分的提取,直到在热点图中找不到相关区域为止。最后,我们将全局特征和最精细部分特征拼接起来,作为支持向量机的输入来进行车型细类识别。整个过程没有加入人类的先验知识,是系统自动学习到了哪些部分有助于识别任务,这体现了我们的系统具有很强的鲁棒性,能够应用到其他相似的识别任务中。利用这个系统,我们在具有281类车型的大规模数据集上获得了 98.29%的精确度,具有了实际的应用价值。更多还原

【Abstract】 Fine-grained vehicle model recognition is an important research topic in the field of intelligent transportation systems.It also belongs to a task in computer vision.Fine-grained vehicle model recognition identifies the specific model of the vehicle from the vehicle picture or video.Different from the traditional vehicle recognition,that is,identifying the types of vehicles,such as cars,trucks,buses,etc.,fine-grained vehicle model recognition will identify a specific model of a specific manufacturer.The research on the traditional vehicle model recognition is mainly carried out through artificial design features and methods based on 3D models.The accuracy and generalization are poor,and it is difficult to apply to the actual intelligent transportation system.In recent years,deep learning,especially convolutional neural network,has a great impact on the artificial intelligence and is currently the dominant method for image recognition.It not only greatly improves the recognition accuracy but also reduces the complexity of the artificial design features.Now,the task of identifying fine-grained vehicle model has gradually begun to use convolutional neural networks.This paper proposes a detailed classification system based on the coarse-to-fine convolutional neural network architecture.In this system,we will use the global features and the finest local features of the models extracted from large-scale datasets for vehicle identification.In order to get the finest local features,we have proposed two important algorithms,one is positioning local area algorithm,and the other is from coarse-to-fine detection algorithm.For the local localization algorithm,we will establish a mapping relationship between the feature heatmap and the input data from the convolutional neural network,so as to extract the most discriminative information part in the input data according to the heat area in the heatmap.For coarse-to-fine detection algorithms,we will repeatedly use the local positioning region algorithm to extract the refined parts again,until no relevant region can be found in the heatmap.Finally,we combine the global features with the finest features,and used as the input of the SVM to identify vehicle models.The entire process does not include prior knowledge of human beings.The system automatically learns which part will help to identify models.This shows that our system is very robust and can be applied to other similar recognition tasks.Using this system,we achieved 98.29%accuracy on a large-scale dataset with 281 models,which have practical application value.更多还原