基于多任务Faster R-CNN车辆假牌套牌的检测方法

针对现有车辆假牌套牌各种检测方法存在计算复杂度高、检测精度低、鲁棒性欠缺等问题,提出一种基于多任务的高速区域卷积神经网络(Faster R-CNN)车辆假牌套牌的检测方法。首先利用时空约束得到疑似套牌车辆,接着用Faster R-CNN定位分割出车辆前脸部分图像,然后对疑似套牌车辆的车脸公脸部分(车辆的基本特征)的特征进行比对;在此基础上再对高仿套牌车辆的车脸私脸部分(车检标)的细微特征进行检测比对。这种分层次的、从车辆宏观特征到微观特征的视觉检测方法,具有检测速度快、鲁棒性高、泛化能力强、实施部署方便、检测精度高等优点。实验研究表明,在Vehicle ID数据集和杭州卡口数据集中分别取得了99.39%、99.22%的检测精度。

Detection method for fake plate vehicles based on multitask Faster R-CNN

Current vehicle detection method for fake plate vehicles has a high computational complexity, low detection accuracy, lack of robustness. This paper presents of fake plate vehicles detection method based on multitask Faster R CNN (region based convolutional neural network). Firstly, spatio temporal constraint is used to obtain the suspected fake plate vehicle. Then, front part of the vehicle is located in the image using Faster R CNN. Next, the public face (basic characteristics of a vehicle) of suspicious fake plate vehicles is contrasted. In further, the subtle features of a private face (Annual inspection certificate for vehicles) is contrasted. This hierarchical visual inspection method, detected from macroscopic features of vehicles to microscopic features, has the advantages of fast detection speed, high robustness, strong generalization ability, convenient deployment and high detection precision. Experimental results show that detection accuracy are 99.39% and 99.22% on the Vehicle ID data set and the Hangzhou bayonet data set, respectively.