License plate detection algorithm based on improved YOLOv8

The issues of challenging license plate localization, sluggish detection, and low accuracy in challenging situations including low resolution, multi-vehicle traffic, and lighting are addressed by a proposed technique to enhance YOLOv8. Three improvements are made to the suggested approach based on the YOLOv8 framework. The license plate’s shape and contour features can be more precisely and effectively recorded by using AKConv in the model’s neck network, which automatically modifies the convolution kernel’s shape based on the object’s size and shape; Second, the SA attention mechanism is added to the C2f module, which betters the accuracy and performance of the system by drawing attention to the license plate’s features from two dimensions. Third, we use SPPE_EMA rather than the conventional SPPF in order to increase prediction accuracy. With these enhancements, the model can more effectively extract features from license plate images, increasing the identification accuracy. Results from experiments on the public dataset CCPD demonstrate that the enhanced Yolov8 performs noticeably better than the original Yolov8. Specifically, assuming that the model’s number of parameters has been decreased by $\mathbf{1 2 \%}$, the precision has increased by 0.9%, and the map50-95 has improved by 2.6%, respectively, our method offers an efficient number plate detection solution.