Convolutional neural networks-based intelligent recognition of Chinese license plates
Convolutional neural networks-based intelligent recognition of Chinese license plates
复制标题
DOI：
10.1007/s00500-017-2503-0
复制
发表时间：
2018-04-01
期刊：
影响因子：
4.1
通讯作者：
Huang, Tingwen
中科院分区：
文献类型：
--
作者：
Liu, Yujie;Huang, He;Huang, Tingwen
车牌识别在许多领域得到了广泛的应用。一些有趣的算法和模型已经被开发来处理定位、分割和识别过程中的问题。本文通过设计新的卷积神经网络，对具有日常生活背景的车牌进行智能识别。首先，针对日常生活背景图像中车牌提取比固定背景图像中车牌提取困难的问题，提出了一种彩色边缘检测算法。针对质量较差的板材样品，提出了一种颜色降低的灰度转换方法，并给出了一种改进的重定位方法，消除了板材边框。然后结合连通域分析和投影分析对图像进行分割。最后，设计了简化的递归卷积神经网络来自动识别字符（第一个是汉字，其次是六个字母数字字符）。人工采集了2189幅不同背景的中国车牌图像。对这些样本进行测试，我们的算法分别达到的定位率、分割率和识别率。该系统对车牌的识别准确率达到了90%以上，完成一个车牌的识别平均需要318 ms，满足实时性要求。
License plate recognition has gained extensive applications in many fields. Some interesting algorithms and models have been developed to deal with the issues in the location, segmentation and recognition processes. This paper focuses on the intelligent recognition of Chinese license plates with daily life backgrounds by designing new convolutional neural networks. Firstly, to extract Chinese license plates from the images subject to daily life backgrounds, which is more difficult than from those with fixed background, a color edge algorithm is proposed to detect specific edges of input image. A color-depressed grayscale conversion method is presented to preprocess plate samples with poor quality, and an improved relocation method is given to eliminate plate frames. Then a combination of connected component analysis and projection analysis is implemented for the segmentation. Finally, simplified and recurrent convolutional neural networks are designed to automatically recognize the characters (the first one is Chinese character, which is followed by six alphanumeric characters). A total of 2189 images containing Chinese license plates are collected manually with different backgrounds. Tested on these samples, the location rate of , segmentation rate of and recognition rate of are, respectively, achieved by our algorithms. The accuracy rate of recognition of Chinese license plates reaches , and it averagely takes 318 ms to complete the recognition of a license plate, which meets the real-time processing requirement.