基于神经网络和支持矢量机的多机动车车牌在线检测方法

Aiming at the problem of multi-license plate recognition in road traffic, a fast and robust multi-license plate detection and recognition method is proposed, which includes two parts: multi-license plate detection and license plate character recognition. A BP (Back-Propagation) neural network model is constructed for color recognition, and combined with the image morphology operation method, candidate license plate targets are screened, and the real license plate targets are discriminated from the candidate license plate targets based on the support vector machine. Through contour size judgment and combined with the license plate size characteristics, the urban code character block, the provincial code character block and the 5-digit motor vehicle coding character block are segmented and extracted in turn, and finally the content of the character block is recognized based on the BP neural network. Based on the above principles, a robust automatic detection and recognition system for multi-motor vehicle license plates has been developed and experimentally tested in real scenes. The results show that: 1) Under the condition that the vehicle is traveling at a normal speed, the system can still ensure a license plate detection and recognition accuracy rate of more than 90%; 2) The system can realize simultaneous multi-license plate detection and recognition; 3) Under the experimental hardware configuration in this paper, the average detection and recognition time of a single image of the system is less than 130 ms, and the processing frequency is about 8 Hz.

针对道路交通多车牌识别问题,提出了一种快速鲁棒的多车牌检测识别方法,包括多车牌检测和车牌字符识别两部分:构造BP (Back-Propagation)神经网络模型用于颜色识别,结合图像形态学运算方法,筛选候选车牌目标,基于支持矢量机从候选车牌目标中判别真正的车牌目标;通过轮廓尺寸判断,并结合车牌尺寸特征,依次分割提取城市代码字符块、省份代码字符块及5位机动车编码字符块,最后基于BP神经网络识别字符块内容.基于上述原理,开发了鲁棒的多机动车车牌自动检测识别系统,并在真实场景中进行了实验测试,结果表明:1)车辆在正常速度行驶条件下,系统依然可以保证90%以上的车牌检测识别正确率;2)系统可实现同时多车牌检测识别;3)文中实验硬件配置下,系统单幅图像检测识别平均时间低于130 ms,处理频率约8 Hz.