|
|
--- |
|
|
license: mit |
|
|
language: |
|
|
- en |
|
|
metrics: |
|
|
- accuracy |
|
|
tags: |
|
|
- CNN |
|
|
- NeuralNetwork |
|
|
--- |
|
|
# CIFAR-10 CNN Image Classifier |
|
|
|
|
|
A Convolutional Neural Network (CNN) built **from scratch** using **TensorFlow/Keras** to classify images from the CIFAR-10 dataset into 10 object categories. |
|
|
|
|
|
This project focuses on **understanding CNN design, training stability, regularization, and evaluation**, without using pretrained models or transfer learning. |
|
|
|
|
|
--- |
|
|
|
|
|
## 🚀 Project Overview |
|
|
|
|
|
This project demonstrates: |
|
|
|
|
|
- CNN architecture design from first principles |
|
|
- Training and evaluation on the CIFAR-10 dataset |
|
|
- Overfitting detection and mitigation |
|
|
- Confusion matrix–based error analysis |
|
|
- Clean, modular ML project structure |
|
|
|
|
|
The goal is to gain **hands-on understanding of deep learning fundamentals**, rather than maximizing benchmark scores. |
|
|
|
|
|
--- |
|
|
|
|
|
## 🧠 Dataset |
|
|
|
|
|
**CIFAR-10** |
|
|
|
|
|
- 60,000 color images (32×32) |
|
|
- 10 classes: |
|
|
- airplane, automobile, bird, cat, deer |
|
|
- dog, frog, horse, ship, truck |
|
|
- 50,000 training images |
|
|
- 10,000 test images |
|
|
|
|
|
--- |
|
|
|
|
|
## 🏗️ Model Architecture |
|
|
|
|
|
- **3 Convolutional blocks** |
|
|
- Conv2D → Batch Normalization → ReLU → MaxPooling |
|
|
- **Classifier** |
|
|
- Dense(256) → Dropout(0.5) |
|
|
- Dense(128) → Dropout(0.3) |
|
|
- Dense(10 logits) |
|
|
- Loss handled using `SparseCategoricalCrossentropy(from_logits=True)` |
|
|
|
|
|
**Total parameters:** ~1.3M |
|
|
|
|
|
--- |
|
|
|
|
|
## ⚙️ Training Strategy |
|
|
|
|
|
- Input normalization |
|
|
- Data augmentation (horizontal flip, rotation, zoom) |
|
|
- Batch Normalization for stable training |
|
|
- Dropout for regularization |
|
|
- Early stopping to prevent overfitting |
|
|
|
|
|
Training was stopped automatically once validation performance stopped improving. |
|
|
|
|
|
--- |
|
|
|
|
|
## 📊 Evaluation & Results |
|
|
|
|
|
- **Best validation accuracy:** ~67% |
|
|
- Small train–validation gap → good generalization |
|
|
- Performance analyzed using a **confusion matrix** |
|
|
|
|
|
Key observations: |
|
|
- Strong performance on classes like automobile, frog, ship, and truck |
|
|
- Expected confusion between visually similar classes (cat ↔ dog, deer ↔ horse) |
|
|
- Confusion matrix used as a diagnostic tool rather than accuracy alone |
|
|
|
|
|
--- |
|
|
|
|
|
## 🛠️ Installation |
|
|
|
|
|
```bash |
|
|
git clone https://github.com/revanthreddy0906/cifar10-cnn-image-classifier.git |
|
|
cd cifar10-cnn-image-classifier |
|
|
pip install -r requirements.txt |
|
|
``` |
|
|
--- |
|
|
|
|
|
**📌 Key Learnings** |
|
|
-------------------- |
|
|
|
|
|
- CNNs outperform dense networks for image data |
|
|
|
|
|
- Correct data pipelines are critical for stable training |
|
|
|
|
|
- Overfitting must be diagnosed using validation metrics |
|
|
|
|
|
- Confusion matrices provide deeper insight than accuracy alone |
|
|
|
|
|
- Regularization and early stopping are essential for generalization |
|
|
|
|
|
* * * * * |
|
|
|
|
|
**📈 Future Improvements** |
|
|
-------------------------- |
|
|
|
|
|
- Stronger data augmentation (MixUp / CutOut) |
|
|
|
|
|
- Learning rate scheduling |
|
|
|
|
|
- Residual connections (ResNet-style blocks) |
|
|
|
|
|
- Transfer learning with pretrained backbones |
|
|
|
|
|
- Per-class precision and recall analysis |
