一种基于Latent SVM的车辆图像分类方法  

A Vehicle Image Classification Method Based on Latent SVM

作　　者：杜小龙 黄树成[1] DU Xiaolong;HUANG Shucheng(School of Computer,Jiangsu University of Science and Technology,Zhenjiang 212003)

机构地区：[1]江苏科技大学计算机学院,镇江212003

出　　处：《计算机与数字工程》2025年第3期803-810,共8页Computer & Digital Engineering

基　　金：国家自然科学基金项目“基于鲁棒表现建模的目标跟踪方法研究”(编号:61772244)资助。

摘　　要：针对城市中大量重型车辆造成交通拥堵以及传统图像分类在特征提取过程中出现的信息丢失,导致分类精度下降的问题,论文提出了一种基于Latent SVM的车辆图像分类方法。通过更细致的车辆图像分类,可以使交通管控者快速定位到重型车辆,使其驶离城市中心,从而让交通得到极大缓解。该方法通过采用一种新的零件定位算法,自动在每类车辆中找到一组有区别的零件,使用这些零件的特征和它们之间的空间关系来训练每类的模型。此外使用多类数据挖掘方法,在训练过程中过滤困难负样本。最后,将这些经过训练的单个模型结合在一起,可以高精度地对车辆品牌和车型进行分类。在CompCars数据集上的实验结果表明,该方法具有令人满意的特征提取能力和更精确的分类能力。In order to solve the problem of traffic congestion caused by a large number of heavy vehicles in the city and the loss of information in the process of feature extraction of traditional image classification,which leads to the decrease of classification accuracy,this paper proposes a vehicle image classification method based on Latent SVM.With more detailed vehicle image classification,traffic controllers can quickly locate heavy vehicles and drive them away from the city center,thereby greatly easing traffic.By adopting a new part localization algorithm,the method automatically finds a set of distinct parts in each class of vehicle,and uses the features of these parts and the spatial relationship between them to train the model of each class.In addition,a multi-class data mining method is used to filter difficult negative samples in the training process.Finally,these trained individual models are combined to classify vehicle brands and models with high accuracy.Experimental results on CompCars datasets show that the proposed method has satisfactory feature extraction ability and more accurate classification ability.

关 键 词：图像分类 零件定位 Latent SVM 特征提取 

分 类 号：TP393.08[自动化与计算机技术—计算机应用技术]