A system to automatically recognize vehicle license plates is a growing need to improve safety and traffic control, specifically in major urban centers. However, the license plate recognition task is generally computationally intensive, where the entire input image frame is scanned, the found plates are segmented, and character recognition is then performed for each segmented character. This paper presents a methodology for engineering a system to detect and recognize Brazilian license plates using convolutional neural networks (CNN) that is suitable for embedded systems. The resulting system detects license plates in the captured image using Tiny YOLOv3 architecture and identifies its characters using a second convolutional network trained on synthetic images and fine-tuned with real license plate images. The proposed architecture has demonstrated to be robust to angle, lightning, and noise variations while requiring a single forward pass for each network, therefore allowing faster processing compared to other deep learning approaches. Our methodology was validated using real license plate images under different environmental conditions reached a detection rate of 99.37% and an overall recognition rate of 98.43% while showing an average time of 2.70 s to process \(1024 \times 768\) images with a single license plate in a Raspberry Pi3 (ARM Cortex-A53 CPU). To improve the recognition accuracy, an ensemble of CNN models was tested instead of a single CNN model, which resulted in an increase in the average processing time to 4.88 s for each image while increasing the recognition rate to 99.53%. Finally, we discuss the impact of using an ensemble of CNNs considering the accuracy-performance trade-off when engineering embedded systems for license plate recognition.

中文翻译：

---

使用深度学习的嵌入式自动车牌识别系统

自动识别车辆牌照的系统对改善安全和交通控制的需求不断增长，特别是在主要城市中心。但是，车牌识别任务通常是计算密集型的，其中扫描整个输入图像帧，对找到的车牌进行分割，然后对每个分割的字符执行字符识别。本文介绍了一种用于卷积神经网络（CNN）的工程系统来检测和识别巴西车牌的方法，该方法适用于嵌入式系统。最终的系统使用Tiny YOLOv3架构检测捕获的图像中的车牌，并使用在合成图像上训练并用实际车牌图像进行微调的第二个卷积网络来识别其字符。所提出的架构已证明对角度，闪电和噪声变化具有鲁棒性，同时每个网络都需要一个前向通过，因此与其他深度学习方法相比，处理速度更快。我们的方法在不同环境条件下使用真实车牌图像进行了验证，检出率为99.37％，总识别率为98.43％，同时平均处理时间为2.70 s在Raspberry Pi3（ARM Cortex-A53 CPU）中具有单个牌照的\（1024 \ times 768 \）图像。为了提高识别精度，测试了CNN模型的集成而不是单个CNN模型，这导致每个图像的平均处理时间增加到4.88 s，同时将识别率提高到99.53％。最后，我们讨论了在设计嵌入式系统进行车牌识别时考虑准确性与性能之间的权衡的情况下使用CNN集成的影响。