Handwritten Digit Recognition using CNN

Project Overview

This project implements a Convolutional Neural Network (CNN) to recognize handwritten digits (0–9) using the MNIST dataset. The model is trained to classify grayscale images of size 28×28 pixels and can also predict custom handwritten digit images.

---

Features

• Load and preprocess MNIST dataset
• Build CNN model using TensorFlow and Keras
• Train and evaluate the model
• Achieve approximately 99% accuracy
• Predict individual test images
• Generate confusion matrix
• Test custom handwritten digit images

---

Dataset

The project uses the MNIST dataset, which contains:

• 60,000 training images
• 10,000 testing images
• 28×28 grayscale images
• 10 classes (digits 0–9)

---

Model Architecture

The CNN model consists of:

• Convolutional Layer (32 filters, 5×5 kernel, ReLU activation)
• MaxPooling Layer
• Dropout Layer (0.2)
• Flatten Layer
• Dense Layer (128 neurons, ReLU activation)
• Output Layer (Softmax activation)

---

Technologies Used

• Python 3.10
• TensorFlow / Keras
• NumPy
• Matplotlib
• Scikit-learn
• OpenCV (for custom image testing)

---

Data Preprocessing

• Reshape images to (28, 28, 1)
• Normalize pixel values from 0–255 to 0–1
• One-hot encode labels
• Resize and center custom images before prediction

---

Model Training

model.fit(X_train, y_train,
          validation_data=(X_test, y_test),
          epochs=10,
          batch_size=200)

---

Model Evaluation

scores = model.evaluate(X_test, y_test)
print("Test Accuracy:", scores[1] * 100)

The model achieves approximately 99% accuracy on test data.

---

Testing Custom Handwritten Image

Steps:

1. Write a digit on white paper
2. Capture and crop the image
3. Convert to grayscale
4. Resize to 28×28
5. Normalize and reshape
6. Predict using:

prediction = model.predict(image)
print("Predicted Digit:", np.argmax(prediction))

---

Confusion Matrix

Used to analyze classification performance:

from sklearn.metrics import confusion_matrix

This helps visualize model accuracy for each digit class.

---

Results

• High accuracy (approximately 99%)
• Good generalization on test data
• Works with properly preprocessed custom images