A New Hybrid Model of Generative Adversarial Network and You Only Look Once Algorithm for Automatic License-Plate Recognition

2025-09-08

AI Summary

To address low ALPR accuracy in complex scenes—particularly under motion blur—this paper proposes an end-to-end trainable hybrid model integrating a selective, lightweight Deblur-GAN with YOLOv5. First, the Deblur-GAN performs targeted deblurring as a preprocessing step to avoid ineffective enhancement; subsequently, license plate detection (LPD), character segmentation (CS), and character recognition (CR) are jointly modeled within the YOLOv5 framework for unified optimization. To support this work, we construct and publicly release the first realistic, motion-blurred ALPR dataset specifically for Iranian license plates. Experiments show that the system processes each frame in just 0.026 seconds, achieving 95% detection accuracy and 97% recognition accuracy. Incorporating the deblurring module improves overall accuracy by nearly 40%, significantly enhancing robustness in blurred scenarios and demonstrating strong potential for embedded deployment.

Technology Category

• Computer Vision: Large Vision Models
• Machine Learning: Hardware-aware ML
• Intelligent Robots: Multimodal Perception & Sensor Fusion

Application Category

• Systems and Infrastructure for Web, Mobile and WoT: Applied ML and AI for Web-based mobile applications
• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking
• Responsible Web: Human-perceived consequences of algorithmic deployment on the web

Abstract

Automatic License-Plate Recognition (ALPR) plays a pivotal role in Intelligent Transportation Systems (ITS) as a fundamental element of Smart Cities. However, due to its high variability, ALPR faces challenging issues more efficiently addressed by deep learning techniques. In this paper, a selective Generative Adversarial Network (GAN) is proposed for deblurring in the preprocessing step, coupled with the state-of-the-art You-Only-Look-Once (YOLO)v5 object detection architectures for License-Plate Detection (LPD), and the integrated Character Segmentation (CS) and Character Recognition (CR) steps. The selective preprocessing bypasses unnecessary and sometimes counter-productive input manipulations, while YOLOv5 LPD/CS+CR delivers high accuracy and low computing cost. As a result, YOLOv5 achieves a detection time of 0.026 seconds for both LP and CR detection stages, facilitating real-time applications with exceptionally rapid responsiveness. Moreover, the proposed model achieves accuracy rates of 95% and 97% in the LPD and CR detection phases, respectively. Furthermore, the inclusion of the Deblur-GAN pre-processor significantly improves detection accuracy by nearly 40%, especially when encountering blurred License Plates (LPs).To train and test the learning components, we generated and publicly released our blur and ALPR datasets (using Iranian license plates as a use-case), which are more representative of close-to-real-life ad-hoc situations. The findings demonstrate that employing the state-of-the-art YOLO model results in excellent overall precision and detection time, making it well-suited for portable applications. Additionally, integrating the Deblur-GAN model as a preliminary processing step enhances the overall effectiveness of our comprehensive model, particularly when confronted with blurred scenes captured by the camera as input.

Problem

• Addressing license plate recognition challenges in blurred conditions
• Improving real-time detection accuracy for intelligent transportation systems
• Reducing computational costs while maintaining high recognition precision

Innovation

• Selective GAN for deblurring preprocessing
• YOLOv5 for license plate detection
• Integrated character segmentation and recognition