Recognizing handwritten digits might sound simple, but it’s actually one of the foundational problems incomputer vision and pattern recognition. Every person writes digits differently — the number “8” might appear looped, stacked, or open-ended depending on handwriting style. The MNIST dataset, a benchmark dataset in AI, captures this diversity. It includes 70,000 grayscale images(60,000 for training and 10,000 for testing), each showing digits from 0 to 9 written by thousands of people.
Traditionally, algorithms like Support Vector Machines (SVM) or Multilayer Perceptrons (MLP)were used to classify such images. However, these models require manual feature extraction — meaning programmers had to decide what visual features (edges, corners, textures) the model should focus on. CNNs completely change that game.

CNNs are designed to automatically detect and learn visual patterns from images through layered processing. Instead of manually defining features, CNNslearn directly from the raw pixel data. This is done through a series of convolutional, pooling, and fully connected layers.

1.Data Preprocessing:

The MNIST images were normalized (values scaled between 0–1) to help the model learn faster.

2. Label Encoding:

The digits (0–9) were converted into a one-hot encoded format, where each number becomes a binary vector.

3. Model Training:

Using TensorFlow and Keras, the CNN was trained for 5 epochs with the Adam optimizer and categorical cross-entropy loss.

4. Evaluation:

The model achieved a 99.31% test accuracy, outperforming traditional models — MLP (96.89%) and SVM (95.31%).

5. Visualization:

The training and validation graphs showed consistent improvement, proving the model was learning effectively without overfitting.

What makes this project unique is itsdeployment as a real-time interactive app Using Streamlit, a Python-based web framework, Ayushmaan integrated the trained model into a web interface. Converts the image to grayscale, Resizes it to 28×28 pixels (the MNIST standard size), Normalizes the input, Passes it through the CNN model for prediction, Within a second, the app displays the predicted digit with high confidence. This live interaction turns a research model into a practical AI tool that can be used for education, document processing, and more

While the CNN model performed excellently, it wasn’t perfect. Certain digits like3 and 8 or 4 and 9 often got confused due to similar shapes. These errors mainly occurred because of variations in writing style, stroke thickness, or incomplete loops in handwritten input.

1. Banking & Finance: Automatically reading cheque amounts or handwritten forms.
2. Education: Interactive learning apps that evaluate handwriting in real time.
3. Healthcare: Digitizing handwritten prescriptions and medical records.
4. Government & Postal Systems: Recognizing postal codes or tax form entries.
5. IoT & Mobile Devices: Deploying lightweight models via TensorFlow Lite for offline recognition.

The study opens doors for more advanced research directions:

1. Transfer Learning: Using pre-trained CNNs to handle more complex data.
2. Hybrid Models: Combining CNN with SVM for better interpretability.
3. Gamified Learning: Making AI-based handwriting trainers for children.
4. Cross-Language Models: Recognizing digits from multiple scripts.

These enhancements can make digit recognition even more intelligent and human-like.

Ayushmaan Singh Yadav’s research clearly demonstrates the power of Convolutional Neural Networks in recognizing handwritten digits with near-human accuracy. With a 99.31% success rate, the model proves that deep learning can outperform traditional techniques while being scalable and practical.

By combining rigorous technical design with real-world usability through a Streamlit web app, this project bridges the gap between theory and application. It shows how artificial intelligence can make data entry faster, reduce human error, and bring automation closer to everyday life.

The future of handwritten digit recognition — and indeed, AI-powered vision systems — is bright, intelligent, and ready to read the world, one digit at a time.

Ayushmaan Singh Yadav

B.tech Computer Science

batch : 2022 - 2026