基于优化卷积神经网络的车辆特征识别算法研究  

Research on vehicle feature recognition algorithm based on optimized convolutional neural network

作　　者：陈暄[1] 吴吉义 CHEN Xuan;WU Jiyi(Zhejiang Industry Polytechnic College,Shaoxing 312000,China;Zhejiang Federation of Artificial Intelligence,Hangzhou 310027,China;Intelligent Education Research Center,Zhejiang University,Hangzhou 310027,China)

机构地区：[1]浙江工业职业技术学院,浙江绍兴312000 [2]浙江省人工智能学会,浙江杭州310027 [3]浙江大学智能教育研究中心,浙江杭州310027

出　　处：《电信科学》2023年第10期101-111,共11页Telecommunications Science

基　　金：国家自然科学基金资助项目(No.61702151,No.61702320,No.61772334);国家重点研发计划项目(No.2018YFB1003800);浙江省哲学社会科学规划课题(No.23NDJC369YB)。

摘　　要：针对道路场景图像中不同距离目标车辆特征存在识别效果弱、精度低的问题,提出一种基于优化卷积神经网络的车辆特征识别算法。首先,采用基于PAN模型的多尺度输入获取不同距离的目标车辆特征;其次,在卷积神经网络结构中加入多池化、BN层和Leaky ReLU激活函数改进网络模型的性能,通过引入混合注意力机制,集中关注车辆图像中的重要特征和区域,从而增强了网络模型的泛化能力;最后,通过构建多层次卷积神经网络结构完成对车辆的特征效果识别。仿真实验结果表明,在单一场景的BIT-Vehicle数据库中,本文算法相比CNN、R-CNN、ABC-CNN、Faster R-CNN、AlexNet、VGG16和YOLOV8在单一目标和多目标识别率方面分别提高了16.75%、10.9%、4%、3.7%、2.46%、1.3%、1%和17.8%、10.5%、2.5%、3.8%、2.7%、1.1%、1.3%,在复杂场景的UA-DETRAC数据库中,本文算法相比其他算法在不同距离目标车辆识别中获得了更加精确的效果。To address the issue of weak identification and low accuracy in recognizing features of target vehicles at different distances in road scene images,a vehicle feature recognition algorithm based on optimized convolutional neural network(CNN)was proposed.Firstly,a multi-scale input based on the PAN model was employed to capture target vehicle features at varying distances.Subsequently,improvements were made to the network model by incorporating multi-pool,batch normalization(BN)layers,and Leaky ReLU activation functions within the CNN architecture.Furthermore,the generalization ability of the network model was enhanced by introducing a hybrid attention mechanism that focuses on important features and regions in the vehicle image.Lastly,a multi-level CNN structure was constructed to achieve feature recognition for vehicles.Simulation experiment results conducted on the BIT-Vehicle database within a single scene show the proposed algorithm’s significant enhancements in single-object and multi-object recognition rates compared to CNN,R-CNN,ABC-CNN,Faster R-CNN,AlexNet,VGG16,and YOLOV8.Specifically,improvements of 16.75%,10.9%,4%,3.7%,2.46%,1.3%,and 1%in single-object recognition,as well as 17.8%,10.5%,2.5%,3.8%,2.7%,1.1%,and 1.3%in multi-object recognition,have been demonstrated by the proposed algorithm,respectively.Over the more complex UA-DETRAC datasets,more precise results have been also achieved by the proposed algorithm in recognizing target vehicles at various distances compared to other algorithms.

关 键 词：车辆识别 卷积神经网络 多尺度输入 

分 类 号：TP391.14[自动化与计算机技术—计算机应用技术]