Scholarlink

<p>Accurate and real-time license plate recognition is essential for intelligent transportation systems, yet traditional Automated License Plate Detection and Recognition (ALPDR) models often struggle under environmental anomalies such as blur, occlusion, and poor lighting. This paper addresses these limitations by proposing a robust anomaly-aware detection framework that integrates a Latent Diffusion Model for data augmentation with a YOLOv12-based architecture enhanced through SimAM and Coordinate Attention mechanisms. The diffusion model generates realistic, anomaly-rich images, enriching the dataset and strengthening model generalization. The YOLOv12 backbone, combined with attention blocks, ensures precise detection by focusing on spatial and contextual features. Grad-CAM visualization is incorporated to interpret model predictions and localize misclassified regions, providing transparency and error insights. Experimental results demonstrate exceptional performance across both the CCPD and Indian Vehicle Datasets, with mAP@0.5 reaching 0.996 and 0.9937, and consistent scores of 0.9108 across mAP@0.5:0.95, along with high precision (0.9831) and recall (0.9747). The suggested model provides better detection accuracy in complicated real-world settings alongside interpretability and resilience, making the system ideal for use in adaptive traffic supervision and control systems monitoring dynamic traffic conditions.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Real-time anomaly reduction in license plate recognition using YOLO with attention mechanism and data augmentation

  • 期刊论文

  • 发表日期：

    2025年10月25日

查看全文
  • Bhawana Srivastava,
  • Poonam Chahal

摘要

Accurate and real-time license plate recognition is essential for intelligent transportation systems, yet traditional Automated License Plate Detection and Recognition (ALPDR) models often struggle under environmental anomalies such as blur, occlusion, and poor lighting. This paper addresses these limitations by proposing a robust anomaly-aware detection framework that integrates a Latent Diffusion Model for data augmentation with a YOLOv12-based architecture enhanced through SimAM and Coordinate Attention mechanisms. The diffusion model generates realistic, anomaly-rich images, enriching the dataset and strengthening model generalization. The YOLOv12 backbone, combined with attention blocks, ensures precise detection by focusing on spatial and contextual features. Grad-CAM visualization is incorporated to interpret model predictions and localize misclassified regions, providing transparency and error insights. Experimental results demonstrate exceptional performance across both the CCPD and Indian Vehicle Datasets, with mAP@0.5 reaching 0.996 and 0.9937, and consistent scores of 0.9108 across mAP@0.5:0.95, along with high precision (0.9831) and recall (0.9747). The suggested model provides better detection accuracy in complicated real-world settings alongside interpretability and resilience, making the system ideal for use in adaptive traffic supervision and control systems monitoring dynamic traffic conditions.