Comparing Different Versions of YOLO for License Plate Detection and Recognition

This paper presents a comparison between different versions of YOLO architecture for license plate detection and recognition. We used these models in two stages. The first stage involves detecting license plates, followed by the second stage, which extracts characters from the identified plates. YOLO versions 4 through 11, including lighter distributions, were trained and tested for both detection and recognition tasks. Precision and recall metrics were used to evaluate the models' performance. For license plate detection, the recall metric was prioritized to minimize false negatives, ensuring all plates were detected. In character recognition, precision became the focus, as misclassification of characters can lead to critical errors in traffic control and law enforcement systems. Additionally, computational efficiency, represented by FLOPs, was considered to reduce costs in real-world applications. We proposed two new metrics combining precision, recall and FLOPS to find which model performs better in both performance and computational cost. The results demonstrate that lightweight models like SATPA and YOLOv7-Tiny are ideal for license plate detection, and similar efficiency can be achieved in character recognition using customized feature extraction networks. The paper concludes by emphasizing the need for tailored models in simpler tasks to achieve a balance between accuracy and computational efficiency.