基于深度学习的动态车牌识别算法研究
Research on Dynamic License Plate Recognition Algorithm Based on Deep Learning
李航行 1阮士峰 1王文童 1张国豪1
作者信息
摘要
[目的]针对当下的车牌识别系统在复杂场景条件下的车牌识别能力相对较弱这一问题,提出了一种基于YOLOv7和LPRNet改进的车牌识别算法.[方法]首先,在YOLOv7头部网络中加入全局注意力模板(GAM),以加强对车牌区域的关注,从而改善车牌的准确定位.其次,利用四点透视变换法对畸形车牌图像进行校正处理,以提供规范化和稳定的输入,有助于后续车牌识别的有效进行.最后,采用LPRNet模型对处理后的车牌图像进行端到端的识别,从而完成对车牌信息的提取和识别.[结果]在CCPD数据集上进行的实验表明,车牌检测精度可达到99%,车牌识别精度可达到99.19%.[结论]研究结果表明,尽管面临复杂场景的挑战,但该车牌识别算法在车牌检测和识别方面仍然保持卓越的效果.
Abstract
[Purposes]In response to the relatively weak license plate recognition capability of current li-cense plate recognition systems under complex scene conditions,an improved recognition algorithm for license plates in complex scenes based on YOLOv7 and LPRNet is proposed.[Methods]First,a global attention template(GAM)is introduced into the YOLOv7 head network to enhance attention to the li-cense plate area,thereby improving the accurate positioning of license plates.Secondly,the four-point perspective transformation method is employed to correct deformed license plate images,providing nor-malized and stable inputs for effective subsequent license plate recognition.Finally,the LPRNet model is applied to achieve end-to-end recognition of the processed license plate images,completing the ex-traction and identification of license plate information.[Findings]Experiments conducted on the CCPD dataset demonstrate a license plate detection accuracy of 99%and a license plate recognition accuracy of 99.19%.[Conclusions]Research results indicate that,despite facing challenges in complex scenes,the proposed license plate recognition algorithm still maintains excellent performance in license plate detec-tion and recognition.
关键词
车牌识别/YOLOv7/注意力机制/透视变换/LPRNetKey words
license plate recognition/YOLOv7/attention mechanism/perspective transformation/LPRNet引用本文复制引用
出版年
2024
NETL
NSTL
万方数据