CNN Image Classification — Mongolian License Plate Length

Overview

This project focuses on classifying Mongolian vehicle license plates based on the number of characters present in the plate.

The task is a binary classification problem:

• Class 0 → Plate contains 6 characters
• Class 1 → Plate contains 7 characters

The model is implemented using PyTorch with a Convolutional Neural Network (CNN) architecture.

---

Problem Statement

Given an image of a vehicle license plate, the goal is to predict whether the plate contains 6 or 7 total characters.

This requires the model to learn visual patterns from plate structures, including variations in formatting and symbols.

---

Dataset

This project uses a synthetic dataset based on Mongolian vehicle license plates.

• Total images: 3000
• Training set: 1000 images
• Test set: 2000 images
• Image type: JPG

Classes

• 0 → Plate contains 6 characters
• 1 → Plate contains 7 characters

Notes

• Some plates include additional symbols such as Soyombo
• Country identifiers like MNG / MGL may appear

Dataset format includes:

• train.csv → training labels
• sample_submission.csv → submission format

---

Model

A Convolutional Neural Network (CNN) is used to extract features from images and perform classification.

Workflow

1. Load dataset from CSV
2. Preprocess images (resize, normalization)
3. Train CNN model
4. Evaluate performance using accuracy
5. Predict on unseen test data
6. Generate submission file

---

Tech Stack

• Python
• PyTorch
• NumPy
• Pandas
• PIL (Image processing)

---

Evaluation

The model is evaluated using Accuracy score.

Submission Format

ID,label
0,0
1,1
2,0

---

How to Run

# Clone repository
git clone https://github.com/BeBecpp/CNN-Image-Classification.git

# Navigate to project
cd CNN-Image-Classification

# Install dependencies
pip install -r requirements.txt

# Train model
python train.py

# Generate submission
python predict.py

---

Results

The model is able to learn meaningful features from plate images and generalize to unseen data.

(Add accuracy here if available, e.g. Accuracy: 0.87)

---

Future Improvements

• Improve model accuracy with hyperparameter tuning
• Add data augmentation
• Try deeper architectures (ResNet, EfficientNet)
• Build a web interface for image upload and prediction

---

Author

Bayarbayasgalan (BeBe)

GitHub: https://github.com/BeBecpp Portfolio: https://nero404.blogspot.com/

---

Note

This project is based on a competition-style dataset and demonstrates practical experience in:

• Deep learning workflows
• Image classification
• Data handling and preprocessing
• Model training and evaluation