Iranian Vehicle Classifier 🚗🇮🇷

This project is developed to classify 20 types of Iranian vehicles, including both trucks and passenger cars, using a custom CNN model built from scratch in PyTorch.

🚀 Project Summary

The goal of this project is to accurately identify close-up images of Iranian vehicles from traffic camera views. Unlike traditional datasets that include full street scenes, our model assumes that the input images are cropped vehicles, extracted from object detection bounding boxes — ensuring focused, high-quality classification.

---

⚠️ Notes:

This model is specifically optimized for cropped vehicle images taken from front view traffic cameras. The model does not perform well on images containing multiple objects, street scenes, or non-vehicle content. Intended use: vehicle classification after detection step. The model is trained on 224x224 RGB images. input images must be Normalized using the following mean and std values.(as used during training):

transform = transforms.Compose([
        transforms.Resize((224, 224)),
        transforms.ToTensor(),

        transforms.Normalize(mean=(0.5,0.5,0.5),
                      std=(0.5,0.5,0.5))
    ])

🌐 Gradio Demo Deployment(Live Test)

The model is deployed using Gradio, allowing users to test it online by uploading an image of a close-up Iranian vehicle.

🚦 Try the model live using the link below:

Live Test

📂 Dataset Description

⚠️ Notes:

Before training, be sure to run augment_and_save() from augmentation.py to generate the augmented dataset The dataset contains a total of 968 images, meaning each class has a limited number of samples.

• ✅ Number of Classes: 20 (Saina,Pride, Pars, Dena, Samand,L90, Renualt-Trucks, Vans, etc.) + Unknown

• 📈 Balanced Dataset: Yes

• 🖼️ Image Type: Close-up vehicle front-view, realistic traffic camera angles

• 🔗 Dataset Link: Dataset

---

🧠 Model Architecture

A custom Convolutional Neural Network (CNN) was designed and trained from scratch in PyTorch with the following features:

• Loss Function : I used CrossEntropyLoss , which is a standard choice for multi-class classification problems.

• 🧩 Skip Connections (inspired by ResNet)

• OPtimizer : The model is optimized using SGD

• 🔁 ReduceLROnPlateau scheduler for automatic LR adjustment

• 🔥 High performance:

  • Train Accuracy: ~99%

  • Test Accuracy: ~99%

  • Both Loss: Extremely low

• ⚡️ Inference speed: ~5.05 ms per 100 images

---

🧪 Training Configuration

• ✅ Designed specifically for Iranian Vehicles

• 🗂️ Trained using custom PyTorch dataloaders

• 🛠️ Augmentation applied using torchvision.transforms

• 📉 Learning rate automatically adjusted when validation loss plateaus

• 🔍 Evaluation includes confusion matrix, accuracy, and loss visualization

---

🖥️ Files Structure:

├── app.py                  # Gradio deployment

├── augmentation.py         # Custom augmentations

├── config.py               # Hyperparameters and paths

├── custom_model.py         # CNN model with skip connections

├── dataset.py              # create train_dataloader & test_dataloader with `from torch.utils.data import DataLoader`

├── main.py                 # Training pipeline

├── Train_Test_loop.py      # Train & test functions

├── model with info.pt      # Trained model

├── notebook.ipynb          # Kaggle training & visualization

├── requirements.txt        # Dependencies

└── README.md

---

Confusion matrix

Sample Prediction

---

🛠 How to Run:

🔗 Clone the Repository

git clone https://github.com/AlirezaChahardoli/Iranian-Vehicles-Classification.git
cd Iranian-Vehicles-Classification

installa Requirements

pip install -r requirements.txt

Run the Gradio app:

python app.py

For training the model from scratch:

python main.py