基于卷积神经网络的车牌识别技术

License Plate Recognition Technology Based on Convolutional Neural Network

为解决字符分割错误造成的车牌识别错误,提出一种基于卷积神经网络的端到端的车牌识别算法,该算法首先采用基于颜色定位、文字定位和边缘检测的方法从自然场景中提取出车牌,由于样本量的问题,采用车牌生成器对车牌样本进行扩充,得到80602张车牌数据,将车牌按照7:1分为训练集和测试集,使用改进的AlexNet网络生成端到端的深度学习模型进行训练,并使用得到的模型进行车牌字符识别,车牌识别准确率达到96.7%。实验结果表明,该方法车牌识别准确率高,且鲁棒性较好。

In this paper, in order to reduce the error of license plate recognition caused by character segmentation mistake, we proposedan end- to- end license plate recognition algorithm based on convolution neural network. Firstly, the algorithm extracts the license plateinformation from a natural setting based on color positioning, character positioning and edge detection. Due to sample size, the license plategenerator is used to enlarge the license plate sample size to obtain the data of 80602 license plates. Next, the license plates are divided intothe training set and the test set by the ratio of 7：1. Then the model is trained using the improved Alex Net network generated end-to-end deeplearning model, and the resultant model is used for license plate character recognition, with accuracy rate reaching 96.7%. At the end, by anempirical test, it is shown that the method has high accuracy and good robustness.