基于深度学习的有遮挡车牌的识别方法研究  

Research on the Recognition Method of Obscured License Plate Based on Deep Learning

作　　者：杨云[1] 王静[1] 姜佳乐 YANG Yun;WANG Jing;JIANG Jiale(School of Electronic Information and Artificial Intelligence,Shaanxi University of Science and Technology,Xi'an Shaanxi 710021)

机构地区：[1]陕西科技大学电子信息与人工智能学院,陕西西安710021

出　　处：《软件》2023年第8期1-8,22,共9页Software

基　　金：国家自然科学基金资助项目(61971272,61601271);国家重点研发重点专项(2019YFC1520204)。

摘　　要：针对遮挡车牌图像中车牌识别存在的车牌号码识别不准确、识别准确率低等问题,提出一种车牌检测网络YOLOv5s_BCG和车牌识别网络CRNN_Den。在车牌检测网络中,使用BiFPN(BI-directional Feature Pyramid Network)替代原YOLOv5网络中的PANet结构,使网络能够自适应学习不同特征层的重要性权重;融入CBAM(Convolutional Block Attention Module)注意力机制,加强网络对关键特征信息的提取;使用Ghost卷积替换传统卷积降低参数量。在车牌识别网络中,用DenseNet模块替换CRNN(Convolutional Recurrent Neural Network)网络中的特征提取部分,实现特征重用,提升网络性能。实验结果表明,车牌检测网络YOLOv5s_BCG相较于原始YOLOv5网络mAP@0.5提升4.2%,可达99.1%。车牌识别网络CRNN_Den能达到94.35%的识别准确率。所提方法能够更加充分地提取车牌图像中的特征信息,可有效识别被遮挡车牌图像中的车牌号码。In order to solve the problems of inaccuracy and low accuracy of license plate recognition in blocking license plate images,a license plate detection network YOLOv5s_BCG and a license plate recognition network CRNN_Den were proposed.In the license plate detection Network,the BiFPN(BI-directional Feature Pyramid Network)was used to replace the PANet structure in the original YOLOv5 network,so that the network can adaptively learn the importance weight of different feature layers,integrated Convolutional Block Attention Module(CBAM)attention mechanism,which strengthened the extraction of key features by the network.Ghost convolution was used instead of traditional convolution to reduce the number of parameters.In the license plate recognition Network,the DenseNet module was used to replace the feature extraction in the Convolutional Recurrent Neural Network(CRNN)to realize feature reuse and improve the network performance.The experimental results show that compared with the original YOLOv5 network mAP@0.5,the license plate detection network YOLOv5s_BCG increases by 4.2%,reaching 99.1%.The license plate recognition network CRNN_Den can achieve 94.35%recognition accuracy.The proposed method can more fully extract the feature information in the license plate image and effectively recognize the license plate number in the blocked license plate image.

关 键 词：车牌号码识别 BiFPN 注意力机制 特征重用 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]