A Delay-Precision-Balanced Approach for License Plate Recognition Based on Fog Computing Paradigm

With the rapid increase in the number of vehicles, research on license plate recognition (LPR) is becoming increasingly important. However, most of the existing research fails to meet the requirement on fast recognition speed and high recognition accuracy concurrently. In this article, a delay-precision-balanced approach for LPR based on fog (DLPRF) computing is proposed, with the aim of reducing the response latency while ensuring the recognition accuracy. Specifically, a new license plate detection (LPD) framework YOLOv5LPD is proposed, which employs the bidirectional feature pyramid network (BiFPN) as an intermediate layer to effectively integrate features of different scales. By means of removing the detection head for large objects, YOLOv5LPD is able to simplify the network structure and reduce the complexity. Additionally, a lightweight convolutional recurrent neural network (CRNN) is presented for license plate character recognition without the operation of character segmentation. Furthermore, a fog computing-based LPR task offloading algorithm is proposed to significantly reduce the queuing delay in task transmission and execution. Experimental results demonstrate that with dataset Chinese City parking dataset (CCPD), YOLOv5LPD achieves an <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">$mAP0.5:0.95$ </tex-math></inline-formula> of 0.786 in LPD, with a speed improvement of 7% and 30% over YOLOv5s and YOLOv4, respectively. In addition, YOLOv5LPD still holds 95.2% accuracy in cross-data LPR experiments, demonstrating the robustness of the model. Meanwhile, the task offloading algorithm in DLPRF also greatly enhances response speed.