YOLO-SLD: An Attention Mechanism-Improved YOLO for License Plate Detection

Abstract

The vehicle license plate detection plays a key role in Intelligent Transportation Systems. Detecting license plates, such as cars, trucks, and vans, is useful for law enforcement, surveillance, and toll booth operations. How to detect license plates quickly and accurately is crucial for license plate recognition. However, the uneven light condition or the oblique shooting angle of the license plate to be detected changes dramatically in real-world complex capture scenarios and the detection difficulty increases. At the same time, distance, lighting, angle, and other requirements are quite high, which seriously affects the detection performance. Therefore, an improved YOLOv7 integrating the parameter-free attention module SimAM for license plate detection was proposed, namely YOLO-SLD. Without modifying the original ELAN architecture, which is the key component of YOLOv7, a SimAM mechanism was added at the end of the ELAN to better extract license plate features and increase computational efficiency. More importantly, the SimAM module does not require any parameters to be added to the original YOLOv7 network, reducing model computation, and simplifying the calculation process. The performance of the detection model with different attention mechanisms was tested on the CCPD dataset for the first time and the proposed method was proven to be effective. The experimental result shows that the YOLO-SLD model has higher detection accuracy and is more lightweight with mAP at 0.5 with the overall improvement in accuracy from 98.44% of the original YOLOv7 model to 98.91%, an increase of 0.47% in accuracy. The accuracy of the CCPD test subset in dark and light images has improved from 93.5% to 96.7%, an increase of 3.2% in accuracy. The parameter size of the model is reduced by 1.2 million parameters compared to the original YOLOv7 model. Its performance is better than the other prevalent license plate detection algorithms.