基于卷积神经网络的模糊车牌图像检测与识别优化  

Optimization of fuzzy license plate image detection and recognition based on convolution neural network

作　　者：于昊生 王素芬[2] 李富有[3] 莫嘉颖 黎烔辉 YU Haosheng;WANG Sufen;LI Fuyou;MO Jiaying;LI Jionghui(School of Public Security Big Data Modern Industry,Guangxi Police College,Nanning 530028,China;Glorious Sun School of Business and Management,Donghua University,Shanghai 200051,China;School of Economics and Finance,Xi'an Jiaotong University,Xi'an 710049,China;Delivery and Service Department of Nanning,KEDACOM,Nanning 530022,China)

机构地区：[1]广西警察学院公安大数据现代产业学院,广西南宁530028 [2]东华大学旭日工商管理学院,上海200051 [3]西安交通大学经济与金融学院,陕西西安710049 [4]苏州科达科技股份有限公司南宁分公司交付与服务部,广西南宁530022

出　　处：《广西大学学报（自然科学版）》2023年第4期985-996,共12页Journal of Guangxi University（Natural Science Edition）

基　　金：国家社会科学基金项目(20ZA051)。

摘　　要：针对当前深度学习网络模型对模糊车牌图像的特征识别能力有限、识别精度较低和速度较慢等问题,提出以二阶优化算法,即以共轭梯度法与拟牛顿法作为卷积神经网络识别模型的优化算法,对模糊车牌图像进行检测和识别,并与一阶优化算法梯度下降算法作为优化算法的模型从识别精确度、识别时间、收敛速度3个方面进行对比。实验结果表明:共轭梯度法与拟牛顿法的识别准确率分别达到了96.9%和96.6%,相比于梯度下降算法的76.1%有所提高,识别时间和收敛速度差距均处于可接受范围内。In view of the problems of the limited feature recognition ability,low recognition accuracy and slow speed of the current deep learning network model for fuzzy license plate images,second-order optimization algorithms,namely conjugate gradient method and quasi-Newton method,were used to optimize the convolution neural network recognition model for fuzzy license plate image detection and recognition.The recognition accuracy,recognition time and convergence speed of the optimization model were compared with first-order optimization algorithm and gradient descent algorithm.The experimental results show that the recognition accuracy of conjugate gradient method and quasi-Newton method is 96.9%and 96.6%respectively,which is higher than that of gradient descent algorithm(76.1%).

关 键 词：卷积神经网络 共轭梯度算法 拟牛顿算法 车牌识别 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]