Computer Engineering and Applications ›› 2021, Vol. 57 ›› Issue (22): 208-214.DOI: 10.3778/j.issn.1002-8331.2012-0022

• Graphics and Image Processing • Previous Articles     Next Articles

Lightweight Multi-scale Attention Fusion Algorithm for License Plate Detection

ZHANG Shihao, YANG Xiujun, WU Linhuang, CHEN Pingping   

  1. College of Physics and Information Engineering, Fuzhou University, Fuzhou 350000, China
  • Online:2021-11-15 Published:2021-11-16



  1. 福州大学 物理与信息工程学院,福州 350000


License plate recognition technology plays an important role in traffic management, among which license plate detection has a significant impact on the subsequent recognition performance. The existing license plate detection system is easy to be interfered by external environment and has poor detection performance in natural scenes. In this paper, a license plate detection network model based on multi-scale attention fusion is proposed. The pyramid network feature map and CBAM(Convolutional Block Attention Module) attention structure are used to improve the detection accuracy of small targets. At the same time, this method can not only accurately detect and locate the license plate under natural scenes, but also accurately locate the four corners of the license plate, which is beneficial to the subsequent application of license plate recognition. In the experiment, the CCPD data set is amplified by data enhancement method, which effectively alleviates the influence of complex environment changes on license plate detection and enhances the robustness of the model. By training and testing the model, the average accuracy rate of 98.05% and recall rate of 98.71% are obtained, which are better than other license plate detection methods. Moreover, the frame rate reaches 64 frame/s and the real-time performance is high.

Key words: deep learning, target detection, license plate detection, attention mechanism


车牌识别技术在交通管理中发挥着重要作用,其中车牌检测环节对后续识别性能有重大影响。现有的车牌检测系统容易受到外部环境的干扰,在自然场景下的检测性能差。提出一种基于多尺度注意力融合的车牌检测网络模型,利用金字塔网络特征图和CBAM(Convolutional Block Attention Module)注意力结构,提高小目标的检测精度。同时该方法不仅能够准确地检测定位出自然场景下的车牌,还能精确地定位出车牌的4个角点,有利于后续的车牌识别应用。实验中采用数据增强方法对CCPD数据集进行扩增,有效缓解了复杂环境变化对车牌检测造成的影响,增强了模型鲁棒性。通过对模型进行训练和测试,获得了98.05%的平均精确率和98.71%的召回率,优于其他车牌检测方法,并且帧率达到64?frame/s,实时性高。

关键词: 深度学习, 目标检测, 车牌检测, 注意力机制