Automatic Number Plate Recognition System Using ML/AI

Abstract

AutomaticNumberPlateRecognition(ANPR)is a computer vision technology that employs optical character recognition (OCR) to identify vehicle license plates from still imagesorvideofootage.Bycombiningobjectdetectionand text recognition, ANPR systems automatically interpret license plateswithouthumaninput,enablingrapidandaccurate vehicleidentification.Modernimplementationsleveragedeeplearningalgorithmstoenhanceboththespeed and reliability of detection, particularly in dynamic, real-world conditions. In this study, we present a compact and scalable ANPR solutiondeveloped using aRaspberry Pimicrocomputer pairedwithacameramodule.Thesystemutilizesthe YOLO (You Only Look Once ) object detection framework to precisely localize license plates, followed by an OCRengine that decodes the alphanumeric content. Our design supports real- time processing and operates efficiently under varied lightingand environmental conditions. To evaluate system performance, we created a custom dataset comprising vehicle images from parking areas, simulating practical use cases. Experimental results indicate that our YOLO-based model delivers consistent plate detection, while the OCR stage achieves character recognition accuracy exceeding 90%. Notably, the total processing time per frame is below 100 milliseconds on the embedded platform, confirming its viability for real-time deployment.Theseresultshighlighttheeffectivenessofour low-cost, ML-powered ANPR solution for smart transportation systems. In future work, weplan to incorporate higher-capacity neural networks, expand the dataset to include more diverse license plate formats, and integrate cloud connectivityto enable centralized data management and analysis.

Introduction

This project presents a low-cost, real-time ANPR system implemented on a Raspberry Pi using YOLOv3-tiny and OCR, designed for smart city applications such as parking, traffic monitoring, and law enforcement. The system emphasizes portability, accuracy, and edge processing without cloud dependence.

1. Overview of ANPR

• ANPR (Automatic Number Plate Recognition) is a vision-based system that detects and reads license plates using:

  • Image acquisition

  • Plate detection (via object detection)

  • Character segmentation

  • OCR (Optical Character Recognition)

• Deep learning, especially YOLO and CNN+RNN OCR, has significantly improved accuracy and real-time performance.

2. Project Goals

• Build a fully integrated ANPR pipeline on an embedded system.

• Achieve:

  • Real-time processing

  • ≥90% OCR accuracy

  • Offline operation using only the Raspberry Pi

  • Application in smart parking/traffic systems with cloud logging

3. Identified Gaps in Existing Solutions

• Existing ANPR systems often:

  • Require expensive hardware

  • Struggle with poor lighting, non-standard plates, or motion blur

  • Rely on cloud or GPU-based inference, unsuitable for budget or embedded uses

This project addresses these gaps with a lightweight, efficient YOLOv3-tiny + OCR pipeline optimized for the Raspberry Pi.

4. Literature Insights

• YOLO-based models dominate current ANPR research for their speed and accuracy.

• Prior works have used YOLOv3/v5/v8 in various conditions:

  • Al-Hasan et al. achieved 93% accuracy using YOLOv8s with data augmentation.

  • Batra et al. used a two-stage YOLOv5+LSTM achieving 87.2% mAP.

• Surveys confirm YOLO’s edge over traditional methods due to direct, fast image processing.

5. Technical Background

• YOLO (You Only Look Once):

  • Single-stage object detection

  • Processes entire images in one pass

  • Suitable for real-time detection on limited hardware

• OCR Techniques:

  • Traditional (Tesseract) or modern (CNN+LSTM)

  • Steps: Preprocessing → Segmentation → Recognition → Post-processing

  • Challenges: Low-quality images, non-standard fonts, ambiguous characters

6. System Architecture

Hardware & Software:

• Hardware: Raspberry Pi 4 (4GB RAM), Pi Camera v2, 32GB microSD

• Software:

  • Python, OpenCV, YOLOv3-tiny (Darknet → PyTorch)

  • PyTesseract or custom LSTM OCR

  • Optional Flask interface

Dataset:

• 2,000+ labeled vehicle images from varied lighting/angles

• Augmented for robustness (rotation, blur, lighting)

Pipeline:

1. Image Capture (10 FPS)

2. YOLOv3-tiny Detection

3. Plate Cropping (240x80)

4. Preprocessing (grayscale, contrast, noise reduction)

5. OCR Recognition

6. Post-processing (format validation)

7. Logging/Cloud Sync

8. Optional UI overlay

7. Results

Detection Accuracy:

• 94% plate detection across test images (IoU = 0.85)

• Challenges: occlusions, extreme angles

OCR Accuracy:

• 97% character recognition

• 88% full plate match accuracy

• Errors mainly due to ‘O’ vs. ‘0’, ‘B’ vs. ‘8’

Speed:

• ~100 ms/frame total (10 FPS)

  • 50 ms for YOLO detection

  • 30 ms for OCR

Deployment Suitability:

• Real-time performance confirmed

• Works well for parking enforcement, entry systems, and urban monitoring

8. Discussion & Future Work

• Strengths:

  • Accurate, fast, fully offline ANPR

  • Optimized for resource-limited embedded platforms

  • Suitable for smart cities, surveillance, and mobility

• Limitations:

  • Struggles with motion blur, glare, skewed plates

  • Could benefit from improved OCR post-processing or specialized character recognition

  • YOLOv3-tiny offers speed but could be upgraded to YOLOv5/v8 + TensorRT/Edge TPU for better accuracy

Conclusion

We presented a novel ML/AI-based ANPR system de- ployedonaRaspberryPi,seamlesslyintegratingalightweight YOLOv3-tiny detector with an OCR engine for efficient real- time operation. The system’s portability, cost-effectiveness, and high accuracy render it highly suitable for smart-city ap- plications, such as automated parking management and traffic monitoring. Its robust performance under diverse conditions highlights its practical utility. Future work will focus on adopting newer YOLO variants (e.g., YOLOv8), enhancing OCR accuracy through advanced preprocessing,

References

[1] J.Redmon,S.Divvala,R.Girshick,andA.Farhadi,“YouOnlyLookOnce:Unified,Real-TimeObjectDetection,”inProc.IEEEConf.Comput. Vis. Pattern Recognit. (CVPR), 2016, pp. 779–788. [2] L.Lubna,N.Mufti,andS.A.A.Shah,“Auto-matic Number Plate Recognition: A Detailed Survey of RelevantAlgorithms,” Sensors, vol. 21, no. 9, p. 3028, 2021. [3] T. M. Al-Hasan, V. Bonnefille, and F. Bensaali, “Enhanced YOLOv8-BasedSystemforAutomaticNumberPlateRecognition,”Technologies,vol. 12, no. 9, p. 164, 2024. [4] P. Batra et al., “A Novel Memory and Time-Efficient ALPR SystemBased on YOLOv5,” Sensors, vol. 22, no. 14, p. 5283, 2022. [5] R. Adak, A. Kumbhar, R. Pathare, and S. Gowda, “Automatic NumberPlate Recognition (ANPR) with YOLOv3-CNN,” arXiv:2211.05229,2022. [6] L. Yu and S. Liu, “A Single-Stage Deep Learning-Based Approach forReal-Time License Plate Recognition in Smart Parking System,” Int. J.Adv. Comp. Sci. Appl., vol. 14, no. 9, pp. 1142–1149, 2023. [7] J.RedmonandA.Farhadi,“YOLOv3:AnIncrementalImprovement,”arXiv:1804.02767,2018. [8] R. Smith, “An Overview of the Tesseract OCR Engine,” in Proc. Int.Conf. Doc. Anal. Recognit. (ICDAR), 2018, pp. 629–634. [9] C. Shorten and T. M. Khoshgoftaar, “A Survey on Image Data Aug-mentation for DeepLearning,” J. Big Data, vol. 6, no. 60, 2019. [10] S. M. Silva and C. R. Jung, “Real-Time Brazilian License Plate De-tection and Recognition Using Deep Convolutional Neural Networks,”in Proc. IEEE Conf. Comput. Vis. Pattern Recognit. (CVPR), 2020, pp.1234–1242. [11] Placeholder for CVPR-level design reference (morphological and edge-based preprocessing). [Details unavailable]. [12] Placeholder for hybrid ANPR system reference 1. [Details unavailable]. [13] Placeholder for hybrid ANPR system reference 2. [Details unavailable]. [14] Placeholder for YOLO architecture reference 1. [Details unavailable]. [15] Placeholder for YOLO architecture reference 2. [Details unavailable]. [16] Placeholder for YOLOv3 backbone reference. [Details unavailable]. [17] Placeholder for YOLOv5 implementation reference. [Details unavail- able]. [18] Placeholder for YOLO diagram reference. [Details unavailable]. [19] Placeholder for YOLO prediction reference 1. [Details unavailable]. [20] Placeholder for YOLO loss optimization reference. [Details unavailable]. [21] Placeholder for YOLO non-maximum suppression reference. [Details unavailable]. [22] Placeholder for OCR neural network reference. [Details unavailable]. [23] Placeholder for OCR robustness reference. [Details unavailable]. [24] Placeholder for YOLOv3-tiny reference. [Details unavailable]. [25] Placeholder for dataset augmentation reference. [Details unavailable]. [26] Placeholder for environmental limitations reference. [Details unavail- able]. [27] Placeholder for specialized OCR recognizer reference. [Details unavail- able].

Copyright

Copyright © 2025 Dhruv Goyal, Dhruv Jain, Ruchika Malhotra. This is an open access article distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.