Bin Zhao, Nan Bu, Jiaming Chang, and Yang Jiang
Feature fusion, license plate recognition, attention mechanism, deep learning, positioning, detection system
This paper proposes a license plate recognition (LPR) system based on the feature fusion strategy, combining fast positioning box YOLO (FPB-YOLO) and dual attention license plate recognition network (DALPRNet). The system enhances FPB-YOLO to efficiently localise license plate regions through an optimised objective function. DALPRNet, an enhanced version of traditional LPRNet, incorporates specialised attention mechanisms and spatial transformation techniques to achieve robust recognition of license plate characters under challenging conditions. Extensive experiments on the extended chinese city parking dataset (CCPD) benchmark demonstrate the system’s exceptional robustness, achieving a license plate localisation precision of 99.3%, character-level accuracy (CA) of 98.85%, plate-level accuracy (PA) of 95.83%, improving upon the baseline LPRNet by 4.66%. On an independent set of 117 complex real-world images, the system attains an end-to-end success rate of 93.2%, outperforming recent state-of-the-art algorithms. The integrated system operates at 72 FPS, confirming its real- time processing capability. This balance of high accuracy and efficient performance validates the system’s effectiveness for practical applications in traffic law enforcement, parking management, and urban security monitoring.
Important Links:
Go Back
