An Efficient and Layout-Independent Automatic License Plate Recognition System Based on the YOLO detector (1909.01754v4)

Abstract: This paper presents an efficient and layout-independent Automatic License Plate Recognition (ALPR) system based on the state-of-the-art YOLO object detector that contains a unified approach for license plate (LP) detection and layout classification to improve the recognition results using post-processing rules. The system is conceived by evaluating and optimizing different models, aiming at achieving the best speed/accuracy trade-off at each stage. The networks are trained using images from several datasets, with the addition of various data augmentation techniques, so that they are robust under different conditions. The proposed system achieved an average end-to-end recognition rate of 96.9% across eight public datasets (from five different regions) used in the experiments, outperforming both previous works and commercial systems in the ChineseLP, OpenALPR-EU, SSIG-SegPlate and UFPR-ALPR datasets. In the other datasets, the proposed approach achieved competitive results to those attained by the baselines. Our system also achieved impressive frames per second (FPS) rates on a high-end GPU, being able to perform in real time even when there are four vehicles in the scene. An additional contribution is that we manually labeled 38,351 bounding boxes on 6,239 images from public datasets and made the annotations publicly available to the research community.

• The paper introduces an efficient, layout-independent Automatic License Plate Recognition (ALPR) system using YOLO for integrated detection and classification.
• The system achieves a high end-to-end recognition rate of 96.9% across eight diverse datasets and processes over 70 frames per second.
• Sophisticated data augmentation, model tuning, and dataset-specific rules are employed, with code and data publicly released to facilitate further research.

The paper introduces an efficient and layout-independent Automatic License Plate Recognition (ALPR) system leveraging a You Only Look Once (YOLO)-based detector. This system integrates license plate (LP) detection and layout classification to enhance recognition accuracy through post-processing rules. The approach assesses and optimizes various models to achieve an optimal balance between speed and accuracy at different stages. The model training incorporates multiple datasets alongside sophisticated data augmentation techniques to bolster robustness across varying conditions.

Key metrics and contributions of the system include:

1. Performance Metrics: The system realizes an average end-to-end recognition rate of 96.9% over eight public datasets spanning five regions, surpassing prior works and commercial alternatives, notably in datasets like ChineseLP, OpenALPR-EU, SSIG-SegPlate, and UFPR-ALPR.
2. End-to-End Pipeline:
   • Vehicular Detection: Utilizing an adapted YOLOv2 model, the system demonstrates near-perfect recall (99.92%) with a high precision rate (98.37%), across various vehicle types and conditions.
   • LP Detection and Layout Classification: A modified Fast-YOLOv2 aids in classifying LPs into distinct classes such as American, Brazilian, Chinese, European, and Taiwanese. This classification hones recognition by applying layout-specific rules.
   • Recognition Stage: The CR-NET model is exploited for simultaneous character recognition, ensuring efficient handling of LPs without necessitating character segmentation. Through dataset-specific heuristic rules, character misclassifications are minimized, significantly benefiting layouts with fixed character positions and counts.
3. Efficiency: The system processes more than 70 frames per second on high-end GPU hardware, suggesting applicability in real-world scenarios, including those with multiple vehicles per frame.
4. Data Augmentation and Model Tuning: Distinct data augmentation strategies bolster training versatility, creating balanced character occurrences and simulating real-world variabilities. Network modifications are made to align with dataset characteristics and to lower computational demands while maintaining recognition efficacy.
5. Public Contributions: The annotations, images, network architectures, and weights used in this work have been made publicly available, fostering future advancements and facilitating fair comparisons with competing approaches.

In summary, the paper illustrates a robust and adaptable ALPR solution, combining advanced detection networks, layout classification, and thorough evaluation across multiple challenging datasets, to propel real-time ALPR applications with optimized performance.