🖼️ Image Recognition Model — CIFAR-10

A CNN-based image classifier that recognizes 10 categories of objects including animals, birds, and vehicles. Built with PyTorch, featuring data augmentation, batch normalization, dropout regularization, and a cosine annealing learning rate scheduler.

---

📌 Overview

This project trains a deep convolutional neural network on the CIFAR-10 dataset to classify images into 10 categories:

Label	Class	Label	Class
0	✈️ Airplane	5	🐕 Dog
1	🚗 Automobile	6	🐸 Frog
2	🐦 Bird	7	🐴 Horse
3	🐱 Cat	8	🚢 Ship
4	🦌 Deer	9	🚛 Truck

---

✨ Features

• ✅ Data Augmentation — Random crop, horizontal flip, color jitter for better generalization
• ✅ Batch Normalization — Stable training across all conv layers
• ✅ Dropout Regularization — Spatial dropout in conv blocks + FC dropout
• ✅ AdamW + Cosine LR Scheduler — Modern optimizer with learning rate annealing
• ✅ Label Smoothing — Reduces overconfidence, improves generalization
• ✅ Early Stopping — Automatically stops when validation accuracy plateaus
• ✅ Best Model Checkpointing — Saves the best-performing weights
• ✅ Confusion Matrix & Classification Report — Per-class performance breakdown
• ✅ Visual Inference — Sample predictions displayed with confidence scores

---

🏗️ Model Architecture

Input: (3, 32, 32)
│
├── Block 1: Conv(3→64) → BN → ReLU → Conv(64→64) → BN → ReLU → MaxPool → Dropout2d
├── Block 2: Conv(64→128) → BN → ReLU → Conv(128→128) → BN → ReLU → MaxPool → Dropout2d
├── Block 3: Conv(128→256) → BN → ReLU → Conv(256→256) → BN → ReLU → MaxPool → Dropout2d
│
└── Classifier: Flatten → FC(4096→1024) → ReLU → Dropout
                         → FC(1024→512) → ReLU → Dropout
                         → FC(512→10)

Total Parameters: ~7.2M

---

📦 Requirements

pip install torch torchvision numpy matplotlib seaborn scikit-learn jupyter

Tested with Python 3.10+, PyTorch 2.x. GPU recommended but not required.

---

🚀 Quickstart

1. Clone the repo:

git clone https://github.com/alihamza701/image-recognition-model.git
cd image-recognition-model

2. Launch the notebook:

jupyter notebook image_recognition_model.ipynb

3. Run all cells. CIFAR-10 downloads automatically (~170 MB). Training takes ~10–15 min on GPU, ~45–60 min on CPU.

---

📊 Results

Metric	Value
Test Accuracy	~82–85%
Best Val Accuracy	~83–86%
Epochs (typical)	~25–30

Results may vary slightly depending on hardware and random seed.

---

📁 Project Structure

image-recognition-model/
│
├── image_recognition_model.ipynb   # Main training notebook
├── best_model.pth                  # Best checkpoint (generated after training)
├── image_recognition_checkpoint.pth # Full deployment checkpoint
├── training_curves.png             # Loss & accuracy plots (generated)
├── confusion_matrix.png            # Confusion matrix (generated)
└── README.md

---

🔮 Inference Example

import torch
from torchvision import transforms
from PIL import Image

# Load checkpoint
checkpoint = torch.load('image_recognition_checkpoint.pth', map_location='cpu')
model.load_state_dict(checkpoint['model_state_dict'])
model.eval()

# Preprocess your image
transform = transforms.Compose([
    transforms.Resize((32, 32)),
    transforms.ToTensor(),
    transforms.Normalize((0.4914, 0.4822, 0.4465), (0.2023, 0.1994, 0.2010))
])

img = Image.open('your_image.jpg')
tensor = transform(img).unsqueeze(0)

with torch.no_grad():
    logits = model(tensor)
    pred = logits.argmax(1).item()

CLASSES = ['airplane','automobile','bird','cat','deer','dog','frog','horse','ship','truck']
print(f'Predicted: {CLASSES[pred]}')

---

🛠️ Possible Improvements

• Replace custom CNN with ResNet-18 via transfer learning for ~93%+ accuracy
• Add Grad-CAM visualizations to see what the model focuses on
• Deploy as a Gradio or Streamlit web app
• Experiment with MixUp or CutMix augmentation

---

👤 Author

Ali Hamza

• GitHub: @alihamza701
• LinkedIn: alihamzarasheed
• Kaggle: alihamzarasheed

---

📄 License

This project is open source and available under the MIT License.