王汝心 马维华

摘  要： 针对路边停车的智能化管理程度较低的问题，以现行标准为依据，对车牌识别面临的问题进行分析。根据所捕获图像的特点，通过高斯滤波、顶帽运算、Otsu算法进行图像预处理，消除噪点，完成背景提取和图像二值化。运用闭运算结合车牌的外形特征完成车牌识别。结合投影法、字符间隔及尺寸的測定和轮廓分析法完成字符分割。最后，提取字符图片的HOG特征，设计了一个基于SVM的字符分类器，实现了一个准确率高、适应性强的车牌识别系统。

关键词： 车牌识别; Otsu; 形态学; 支持向量机

中图分类号：TP36          文献标识码：A 文章编号：1006-8228（2021）07-01-05

License plate recognition method based on HOG features

Wang Ruxin， Ma Weihua

（Nanjing University of Aeronautics and Astronautics， Nanjing， Jiangsu 210016， China）

Abstract： In response to the low level of intelligent management of on-street parking， the problems faced by license plate recognition are analyzed based on the current standards. According to the characteristics of the captured image， image pre-processing is performed by Gaussian filtering， Top Hat operation and Otsu algorithm to eliminate noise， and background extraction and image binarization are completed. The license plate recognition is completed by using the closed operation combined with the shape features of the license plate. The character segmentation is completed by combining the projection method， the determination of character interval and size， and the contour analysis method. Finally， HOG features of character images are extracted and a character classifier based on SVM is designed to realize a license plate recognition system with high accuracy and high adaptability.

Key words： license plate recognition; Otsu; morphology; SVM

0 引言

车牌识别（License Plate Recognition，LPR）在交通执法、停车管理、自动收费站等多种场合起着重要的作用，有助于智能交通系统的发展。车牌识别主要包括三个部分：在捕获的图像中确定车牌位置，在车牌区域分割字符和识别字符。为了减少环境因素对车牌识别的干扰，应先将图像进行预处理，再进行车牌定位。

常用车牌定位算法主要有三类：通过汽车号牌的底色定位车牌位置;基于滤波器过滤图像中边缘密集的区域;提取车牌特征，结合机器学习算法。常用的字符识别方法有模板匹配、投影分析、结构特征法等[1]。当图片背景颜色较为复杂时，基于彩色信息进行车牌定位需要较大的计算量，而且难以对各种图像做统一处理;神经网络通常算法复杂;边缘检测方法需要设计适合字符的边缘检测算子[2]，但得到的结果难以突出车牌区域。……