Abstract
Automatic License Plate Recognition (ALPR) in regions with mixed plate styles, such as the Pearl River Delta, presents unique challenges for traditional two-stage systems. We present ALPRNet, a novel one-stage neural network that performs license plate detection and character recognition simultaneously. By treating characters as individual objects rather than sequences, ALPRNet eliminates the need for Recurrent Neural Networks (RNNs) and avoids error accumulation inherent in multi-stage pipelines. Our approach achieves state-of-the-art results on multi-style datasets, including a 98.21% end-to-end recognition accuracy on the HZM dataset.
Introduction
Most ALPR systems are optimized for a single plate style. However, in regions like the Hong Kong-Macao-Mainland China border, vehicles often display multiple plates with varying layouts and character sets. Traditional "detect-then-recognize" pipelines typically use a Convolutional Neural Network (CNN) for detection followed by an RNN for Optical Character Recognition (OCR). This sequential approach is prone to error accumulation—where detection failures cascade into recognition errors—and suffers from the high computational cost of RNNs.
ALPRNet addresses these issues by unifying detection and recognition into a single pass. By redefining character recognition as a multi-class object detection task, the system simplifies the architecture and improves robustness across diverse plate styles and multi-line layouts.
Methodology
The ALPRNet architecture is built upon a ResNet-50 backbone integrated with a Feature Pyramid Network (FPN). The system leverages the multi-scale nature of the FPN to handle objects of different sizes: the high-resolution P1 layer is dedicated to small character detection, while the P2 layer handles the larger license plate (LP) bounding boxes.
The core components include:
- Parallel FCOS Detectors: Two anchor-free Fully Convolutional One-Stage (FCOS) detectors operate in parallel. One branch identifies the LP type and location, while the other detects and classifies individual characters across 73 classes (including English, numeric, and Chinese characters).
- Anchor-Free Design: By predicting bounding boxes directly from pixel locations, the model avoids the limitations of pre-defined anchor boxes, making it more adaptable to varying aspect ratios.
- Assembly Module: A lightweight module groups the "bag of characters" into coherent strings based on spatial coordinates and Intersection over Union (IoU), effectively handling multi-line plates.
- Optimized Loss Functions: Training is guided by Dice coefficient loss for class imbalance, L2 loss for "center-ness," and Distance-IoU (DIoU) loss for precise regression.
Experimental Results
The performance of ALPRNet was validated across three diverse datasets:
- HZM Multi-Style Dataset: Featuring vehicles with up to three different plate styles from the Hong Kong-Zhuhai-Macao Bridge. ALPRNet achieved 100% LP detection accuracy and 98.21% end-to-end recognition accuracy. In a "passing rate" test, it reached 99.43%.
- AOLP (Application-Oriented License Plate): On Taiwan license plates, the model surpassed previous state-of-the-art benchmarks, notably improving performance by 3% in the challenging "Road Patrol" category.
- PKU Dataset: For Mainland China plates, the system reached an average detection performance of 99.83%, outperforming YOLO and Faster R-CNN-based alternatives.
Discussion and Conclusion
The primary innovation of ALPRNet is the shift from sequential decoding to parallel object detection. Treating characters as objects allows the network to share context between plate and character features, leading to more robust performance. The inclusion of a "center-ness" branch further refines results by filtering out low-quality predictions far from the object center.
While highly effective, the model requires dense, character-level bounding box annotations for training. Future work may focus on developing more differentiable assembly logic or testing the system's resilience under extreme environmental conditions like severe weather or motion blur. Nonetheless, ALPRNet represents a significant step toward efficient, unified, and multi-style ALPR systems.
Summary updated