Abstract: Traditional methods of license plate location based on edge, color, texture and machine learning need complex feature extraction for license plate image. During the extraction, the training process may result in over-fitting or dimensionality?curse, in addition, recognition result is likely affected by factors like illumination, road environment, image quality, etc. Although the rate of leak recognition is low, the error rate of recognition is high. According to convolutional neural network, a deep learning algorithm can classify the license plate to avoid complex image pre-processing, which is necessary for traditional pattern classification algorithms. It also reduces the reliance on abundant experience of designing feature extraction algorithm. Convolutional neural network has better performance in license plate classification in comparison with BP neural network and support vector machine. Recognition accuracy is as high as 98.25%, which proves that deep learning algorithms have great application prospect in intelligent transportation field.

Key words: license plate classification, convolutional neural network, convolution layer, downsampling layer

摘要： 传统的基于边缘、颜色、纹理及机器学习等方法进行的车牌定位，需要对车牌图像进行复杂的特征提取，不但训练过程容易造成过拟合或者维数灾难，而且识别结果也易受光照、道路环境及图像质量等因素的影响，虽然漏识别率低，但误识别率高。针对车牌分类问题，利用深度学习中的卷积神经网络，避免了传统模式分类算法在前期对图像复杂的预处理，降低了设计提取特征算法时对丰富经验的依赖。综合对比了BP神经网络、支持向量机、卷积神经网络三种算法，实验结果表明，卷积神经网络在车牌分类中具有较好的表现，识别率高达98.25%，也证明了深度学习在智能交通领域具有较大的应用前景。

关键词: 车牌分类, 卷积神经网络, 卷积层, 降采样层