Bangladeshi License Plate Detection & Recognition 🚘

Automatic License Plate Recognition (ALPR) system tailored for Bangladeshi vehicles.
It combines YOLOv8 for plate detection with a hybrid recognition pipeline using EfficientNetB0 and EasyOCR, supported by preprocessing and confidence-based voting mechanisms.

---

📖 Overview

License plate recognition is a crucial component in traffic management, automated toll collection, parking systems, and vehicle surveillance.
Unlike English or numeric plates, Bangla license plates present unique challenges due to:

• Complex Bangla script morphology (consonants, vowels, compound characters).
• Inconsistent fonts and plate designs.
• Lighting variations, noise, and occlusion in real-world scenarios.

This project addresses these challenges through a robust, deep-learning-based pipeline.

---

🏗️ System Architecture

1. Input Image → Raw vehicle image.
2. Detection Module → YOLOv8 detects the license plate region.
3. Plate Extraction → Cropped region passed to preprocessing.
4. Preprocessing Pipelines → Thresholding (Otsu, Adaptive Gaussian, Adaptive Mean).
5. Dual Recognition →
   • EfficientNetB0 CNN (character recognition).
   • EasyOCR (end-to-end text extraction).
6. Confidence-Based Voting → Final license plate prediction.

🛠️ Methodology

🔍 Detection (YOLOv8)

• Small variant (yolov8s) trained on 785 annotated images.
• Achieved 99% detection accuracy.

🧠 Recognition (EfficientNetB0 + EasyOCR)

• EfficientNetB0 trained on ~17,000 cropped character images across 29 Bangla classes.
• EasyOCR used as a complementary OCR method for robustness.
• Final results fused using confidence-based voting.

🖼️ Preprocessing

• Image Enhancement (unsharp masking, bilateral filtering).
• Grayscale & contrast adjustment (top-hat, black-hat filtering).
• Multi-thresholding (Otsu, Adaptive Gaussian, Adaptive Mean).
• Noise removal & morphological operations.
• Character segmentation & resizing (64×64).

---

📊 Results

Tested using 722 unseen vehicle images.

• YOLOv8 Detection:

  • mAP@0.5: 98.79%
  • Precision: 96.13%
  • Recall: 98.73%

• EfficientNetB0 CNN:

  • Character classification accuracy: 99%
  • Full plate recognition: 73.84%

• EasyOCR:

  • Full plate recognition: 74.79%

• Hybrid Ensemble (CNN + EasyOCR):

  • Full plate recognition: 94.90%

---

Full working

1. Detection:

   

2. Cropping using detected bounding box:

   

3. Preprocessing steps for EfficientNet:

   

4. Contour detection and classification:

   

5. Preprocessing and Recongition for EasyOCR

   

6. Ensemble voting between EfficientNetB0 and EasyOCR

7. Final Text Extraction:

   

📂 Dataset

The system was trained and tested using publicly available Bangladeshi license plate datasets:

• Bangladeshi Vehicle License Plate (Kaggle)
• Bangla License Plate Dataset with Annotations
• Bangladeshi Bus & Truck Plates

---

🔮 Future Work

• Expand dataset with more diverse conditions (rain, night, low quality).
• Adaptive preprocessing selection based on input quality.
• Transformer-based OCR to directly predict full plates (reducing segmentation dependency).

---

🙌 Acknowledgments

• Kaggle contributors for open datasets.
• Ultralytics YOLO & EasyOCR community.