Abstract: Aiming at the problems of low efficiency, poor robustness and low recognition precision of existing license plate recognition systems, this paper proposed an end-to-end deep learning model for license plate detection and recognition in a high-precision real-time environment. First, adds an improved channel attention mechanism to the down-sampling process of the YOLOv5 network layer, which incorporates location information to reduce the information loss caused by sampling and improve the feature extraction capability of the model. Secondly, uses the combination of LSTM+ CTC to construct the recognition network to complete the recognition of license plate without character segmentation, which greatly reduces the training cycle and improves the recognition accuracy and efficiency of the model. This paper has conducted extensive experiments on chinese city parking dataset (CCPD), and the results show that the average recognition accuracy of the improved model of license plate recognition proposed in this paper reaches 98.03%, which is significantly better than the traditional license plate recognition technology, and the recognition effect is good in complex environment with strong robustness.

Key words: Deep learning; Target detection; License plate recognition; YOLOv5; Attention mechanism; Long short-term memory

摘要： 针对现有车牌识别技术效率低、鲁棒性差、识别精度不高等问题，文章提出了一种高 精度实时环境下车牌检测和识别的端到端深度学习模型。首先，在 YOLOv5 网络层的下采样 过程中加入了改进的通道注意力机制，该机制加入了位置信息，减少了采样带来的信息损失， 提高了模型的特征提取能力；其次，基于 LSTM + CTC 的组合构建识别网络，完成车牌的无 字符分割识别工作，大大减少了训练周期，提高了模型的识别精度和效率。文章在中国城市 停车数据集（CCPD）上进行了大量实验，结果表明，文中提出的车牌识别改进模型平均识别 精度达到了 98.03%，明显优于传统的车牌识别技术，且在复杂环境识别效果良好，具有较强 的鲁棒性。

关键词: 深度学习；目标检测；车牌识别；YOLOv5；注意力机制；长短期记忆网络