【作者】 李华；

【导师】 余烨；

【作者基本信息】 合肥工业大学 ， 计算机应用技术， 2021， 硕士

【摘要】 作为车辆的关键标志,车标不易被篡改,对车辆信息的提取有着重要的辅助作用。传统基于手工描述子的车标检测算法存在精度低、速度慢等问题,而基于深度学习的检测算法则难以平衡速度和精度。本文基于卷积神经网络(convolutional neural network,CNN)对车标检测进行研究,在现有目标检测算法的基础上,进行了改进和优化,提出了两种车标检测算法:基于YOLOv3的快速车标检测算法YVLDet和基于Center Net的改进车标检测算法CVLDet。基于卡口图像构建的数据集开展实验,实验结果表明,本文提出的车标检测算法在速度和精度上都有一定的提升,更加适用于实际场景中的车标检测任务。本文的主要工作包括以下几个方面:(1)对现有的车标检测方法进行总结和分析:分析了基于传统特征描述子和基于CNN的目标检测方法的优缺点及其在车标检测上的应用,并对基于CNN的目标检测方法中涉及的数据扩充、骨干网络和特征金字塔等方法进行总结。(2)提出一种基于YOLOv3的快速车标检测算法YVLDet:该算法使用深度可分离卷积、空洞卷积以及通道注意力机制构造速度更快的骨干网络结构,结合可变形卷积以及转置卷积设计更有效的上采样策略,使用骨干网络中更浅层特征加强检测分支的空间特征,让检测的特征图包含丰富的空间特征信息。最后结合YOLOv3的检测模块,使用两个不同尺度的检测分支进行检测。实验结果表明,该方法可以在保持精度的同时大大提升了检测的速度。(3)提出一种基于CenterNet的改进车标检测算法CVLDet:首先针对数据集中车标数目少的缺点,提出一种高效的数据扩充方法以提高网络的学习能力。其次,拓展Res Net-18的网络宽度,融合浅层次网络特征以增加检测分支的空间信息。最后将输入的图像融入深层次网络,减少深层次网络特征的抽象性,使检测特征图的空间信息更加丰富。实验结果表明,该方法可以在保证车标检测速度的同时提高准确率。更多还原

【Abstract】 As a key sign of the vehicle,vehicle logos are not easily modified,which play important auxiliary roles in the extraction of vehicle information.Traditional vehicle logo detection algorithms based on handcrafted descriptors mostly have low accuracy and slow speed.Most object detection algorithms based on deep learning cannot maintain a balance between speed and accuracy.Based on the convolutional neural network(CNN),we carry out research on the vehicle logo detection,and two improved algorithms are proposed: a fast vehicle logo detection algorithm based on YOLOv3 and a Center Net-based vehicle logo detection algorithm.Our experiments are carried out on the dataset constructed from images captured on the highway.The experimental results show that our algorithms proposed in this paper have certain improvements in speed and accuracy,which make them more suitable for practical monitoring scenarios.The main works of this thesis are listed as follows:(1)Existing vehicle logo detection methods are summarized and analyzed.The object detection algorithms based on traditional descriptors and CNNs are summarized,and their application in vehicle logo detection are analyzed.Also,some common methods such as data augmentation,backbone network and feature pyramid involved in object detection are summarized.(2)An improved vehicle logo detection algorithm YVLDet based on YOLOv3 is proposed.The algorithm uses depthwise convolution,dilated convolution and channel attention module to construct a faster backbone network;combined with deformable convolution and transposed convolution,it designs a more effective up sampling strategy;it use shallower features in the backbone network to enhance the spatial feature of the detection branches,so that the detected feature map contains rich spatial feature information;finally,combining the detection module of YOLOv3,two detection branches of different scales are used for detection.The experimental results show that the algorithm can improve the detection speed while ensuring the detection accuracy.(3)A Center Net-based vehicle logo detection algorithm CVLDet is proposed.First,aiming at the shortcoming of small number of vehicle logos in the dataset,a data augmentation method is proposed to improve the learning ability of the network.Then,it expands the network width of Res Net-18 and fuses shallow network features to increase the spatial information of detection branches.Finally,it fuses input image feature and deep network feature to reduce the abstraction of the deep network feature,and make the spatial information of the detection feature map richer.Experimental results show that this method can improve the accuracy while ensuring the speed of vehicle logo detection.更多还原