Scholarlink

<p>Automatic detection of ID cards is crucial for banking, e-governance, and digital onboarding, where efficiency, speed, and accuracy are essential. Current object detection models (YOLOv8, YOLOv12n, YOLOv13, EfficientDet, Cascade R-CNN) fall short in dealing with varied ID card formats, computational overheads, and real-time performance on edge devices. This paper presents IDCard-YOLO, a lightweight deep learning structure for effective ID card detection and classification. IDCard-YOLO cooperatively combines a Bidirectional Feature Pyramid Network (BiFPN) for stable multi-scale feature fusion, MobileNet-V2 as a lightweight backbone to lower computational complexity, and Distribution Focal Loss (DFL) to optimise bounding box regression for small text. IDCard-YOLO has been trained and tested on an in-house dataset of Indian government ID cards like Aadhar Card and PAN Card, and the model has been evaluated using the MIDV-500 public benchmark dataset. Statistically robust experiments show that IDCard-YOLO attains 93.1% mean Average Precision (mAP@0.5), which is better than YOLOv8 (91.2%), EfficientDet (88.4%), Cascade R-CNN (90.1%), and YOLO-NAS (89.8%) in custom dataset. The model is highly efficient, using only 0.9G FLOPs (97% less than YOLOv8), with a model size of 6.2&#xa0;MB and real-time processing at 165 FPS on a GPU. Ablation and robustness tests verify strong generalisation in low light, occlusion, rotations, and blur. With its light architecture, high accuracy, and real-time data processing, the proposed model possesses a scalable and efficient solution for automatic identity recognition. IDCard-YOLO provides scalable, accurate, and resource-saving automated ID verification for mobile, IoT, and security use cases, significantly improving identity document recognition.</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

IDCard-YOLO: a lightweight and high-performance deep learning model for real-time ID card detection and classification

  • 期刊论文

  • 发表日期：

    2026年7月27日

查看全文
  • Pradyumn Pandey,
  • Shrabanti Mandal

摘要

Automatic detection of ID cards is crucial for banking, e-governance, and digital onboarding, where efficiency, speed, and accuracy are essential. Current object detection models (YOLOv8, YOLOv12n, YOLOv13, EfficientDet, Cascade R-CNN) fall short in dealing with varied ID card formats, computational overheads, and real-time performance on edge devices. This paper presents IDCard-YOLO, a lightweight deep learning structure for effective ID card detection and classification. IDCard-YOLO cooperatively combines a Bidirectional Feature Pyramid Network (BiFPN) for stable multi-scale feature fusion, MobileNet-V2 as a lightweight backbone to lower computational complexity, and Distribution Focal Loss (DFL) to optimise bounding box regression for small text. IDCard-YOLO has been trained and tested on an in-house dataset of Indian government ID cards like Aadhar Card and PAN Card, and the model has been evaluated using the MIDV-500 public benchmark dataset. Statistically robust experiments show that IDCard-YOLO attains 93.1% mean Average Precision (mAP@0.5), which is better than YOLOv8 (91.2%), EfficientDet (88.4%), Cascade R-CNN (90.1%), and YOLO-NAS (89.8%) in custom dataset. The model is highly efficient, using only 0.9G FLOPs (97% less than YOLOv8), with a model size of 6.2 MB and real-time processing at 165 FPS on a GPU. Ablation and robustness tests verify strong generalisation in low light, occlusion, rotations, and blur. With its light architecture, high accuracy, and real-time data processing, the proposed model possesses a scalable and efficient solution for automatic identity recognition. IDCard-YOLO provides scalable, accurate, and resource-saving automated ID verification for mobile, IoT, and security use cases, significantly improving identity document recognition.