轻量级车牌识别卷积网络

Lightweight Convolutional Network for License Plate Recognition

为了提高车牌识别的准确性,提出一种轻量级车牌识别神经网络。车牌定位阶段,构造了深度为9的卷积神经网络(Convolutional Neural Networks,CNN),首先利用图像预处理与阈值分割融合的方式对车牌进行粗定位,然后对CNN网络进行模型训练,得到网络权重,最后将车牌候选区域输入到CNN模型来实现精准定位车牌。车牌识别阶段,构造了深度为11的CNN网络,首先对准确定位的车牌进行字符分割,并对分割后的字符进行归一化处理,然后将分割后的单个字符输入到CNN模型,实现对字符的识别,最后输出字符识别结果。通过实验验证,所搭建的两个CNN网络能够有效提升车牌的检测和识别准确率。

To improve the accuracy of license plate recognition,a light-weight neural network for license plate recognition was proposed.In the stage of license plate location,a Convolutional Neural network(CNN)with depth of 9 is constructed.Firstly,the license plate was coarsely located by the method of image preprocessing and threshold segmentation and fusion,and then the model training of CNN network was carried out to obtain the network weight.Finally,the license plate candidate region was input into the CNN model to achieve accurate license plate location..In the stage of license plate recognition,a CNN network with depth of 11 was constructed.Firstly,characters of precisely positioned license plates were segmented,and the characters after segmentation were normalized.Then,single characters after segmentation were input into the CNN model to realize character recognition.Experimental results showed that the two CNN networks could effectively improve the detection and recognition accuracy of license plates.