A Lightweight License Plate Recognition Method Based on YOLOv8 (www.mdpi.com)

A new lightweight license plate recognition system leveraging an improved YOLOv8n architecture combined with an enhanced LPRNet model has been proposed to address challenges in intelligent transportation systems, especially for deployment on resource-constrained edge devices. The system tackles limitations of existing methods including large model sizes, slow inference, and low robustness under complex environmental conditions or tilted license plates. By integrating a novel GCE module—splitting feature maps for richer contextual learning followed by squeeze-and-excitation attention—and replacing YOLOv8’s PANet neck with a Bidirectional Feature Pyramid Network (BiFPN), the model achieves efficient multi-scale feature fusion, markedly enhancing small target detection like license plates. For recognition, the authors optimize LPRNet by removing dropout layers in favor of batch normalization to stabilize training and boost generalization, alongside integrating an Efficient Multi-scale Attention module that improves discriminative feature extraction. In addition, the detection network predicts precise four-point coordinates of license plates rather than just bounding boxes, enabling perspective correction of tilted plates prior to recognition, thereby improving accuracy. This cascaded approach—GCE-enhanced YOLOv8n for localization and refined LPRNet for character recognition—delivers a compact yet robust solution suitable for embedded systems and smart city applications, advancing real-time, high-accuracy license plate recognition in diverse and challenging scenarios.