基于集成方法的自动车型识别方法设计  

Design of Automatic Model Recognition Method Based on Integration Method

作　　者：向冲 廖骏杰[2] XIANG Chong;LIAO Junjie(School of Data and Information,Changjiang Polytechnic,Wuhan 430070,China;School of Electronic and Information Engineering,Wuhan Vocational College,Wuhan 430074,China)

机构地区：[1]长江职业学院数据信息学院,湖北武汉430070 [2]武汉职业技术学院电子信息工程学院,湖北武汉430074

出　　处：《自动化仪表》2023年第1期49-54,共6页Process Automation Instrumentation

摘　　要：为了解决单一光学传感器和单一卷积神经网络(CNN)在车型识别精度方面的问题,以及避免后续因识别误差而造成的不必要经济损失,提出了基于CNN与光学传感器融合的收费站车型识别方法。首先,通过1组CNN得到分类概率,并采用光学传感器得到离散的车辆标签。然后,利用梯度提升方法对上述2个方案得到的结果进行集成,以得到最终的车型类别。在大型的车型识别数据集上的试验结果表明,所提出的方法不仅优于传统的车型识别方法,还优于单一的CNN方法和单一的光学传感器方法。In order to solve the problems of single optical sensor and single convolutional neural network(CNN) in car model recognition accuracy, as well as to avoid the subsequent unnecessary economic losses caused by recognition errors, a method of toll station car model recognition based on the fusion of CNN and optical sensor is proposed. Firstly, the classification probability is obtained by a set of CNN, and the discrete vehicle labels are obtained by using optical sensors. Then, the results obtained from the above two schemes are integrated using a gradient boosting method to obtain the final car model categories. Experimental results on a large model recognition dataset show that the proposed method outperforms not only traditional model recognition methods, but also a single CNN method and a single optical sensor method.

关 键 词：深度学习 卷积神经网络 光学传感器 车型识别 数据标注 损失函数 

分 类 号：TH39[机械工程—机械制造及自动化]