Improved LKM-YOLOv10 Vehicle Licence Plate Recognition Detection System Based on YOLOv10

Vehicle transportation plays a crucial role in society. To distinguish between the numerous vehicles on the road, each automobile must have a license plate. As computer vision technology has progressed, relying on humans to manually identify vehicle plates has become increasingly inadequate for handling growing transportation demands. The integration of computer vision significantly boosts the efficiency of recognizing vehicle license plates. This paper presents an enhanced version of the existing YOLOv10 model called LKM-YOLOv10. It incorporates the LDConv module to apply linear deformable convolution, improving the convolutional neural network's performance; it also utilizes the KAN module to introduce a learnable activation function at various nodes, allowing weight parameters to adjust dynamically; additionally, the MDCR module merges multiple depthwise separable convolution layers with varying dilation rates to improve feature representation. The LKM-YOLOv10 model enhances training efficiency for vehicle license plate recognition tasks. After conducting fifty rounds of experiments, this model demonstrates a significant increase in mAP50-95(B) values-showing an 8.5% improvement by round twenty and 2% by round fifty compared to previous models. When using optimal training results from these fifty rounds for inference purposes, LKM-YOLOv10 achieves 51 more effective identifications than its counterpart YOLOv10n, reflecting a percentage increase of 3.9%.