A PROJECT ON –
HANDWRITTEN DIGIT RECOGNITION (MNIST)
SHIVAM SANTOSH (22BCS80019)
SHIVANI (22BCS80190)
ARINDAM ROY (22BCS80139)
ANJALI RAI (22BCS80103)
SUMIT KR CHOUDHARY (22BCS80051)
PROBLEM DEFINITION
• The "Handwritten Digit Recognition Using MNIST Dataset" project involves building a robust and
efficient machine learning model capable of recognizing and classifying handwritten digits ranging from 0
to 9. The project leverages the MNIST dataset, which consists of 60,000 training images and 10,000 test
images of handwritten digits.
• The primary goal is to pre-process the data, build, train, and optimize a neural network model (such as
Convolutional Neural Network - CNN) that can achieve high accuracy in digit classification. The project
explores various techniques including data augmentation, model tuning, and regularization to enhance the
model's performance. The final model will be evaluated based on its accuracy, precision, recall, and F1-
score, and the results will be compared against standard benchmarks.
PROJECT IDENTIFICATION
• Project Title: Handwritten Digit Recognition Using MNIST Dataset
• Project Domain: Machine Learning / Deep Learning
• Project Duration: 4 Months
• Project Supervisor: Prof Prabhjot kaur
• Team Members: Anjali, Shivani, Shivam, Sumit, Arindam
• Tools & Technologies: Python, MNIST Dataset, CNN, TensorFlow/Keras, NumPy, Pandas, Matplotlib
• Objective: To develop a machine learning model that can accurately recognize and classify handwritten
digits from the MNIST dataset.
METHODOLOGY
Data Preprocessing: Each image was normalized to ensure uniformity and improve model performance.
We also employed data augmentation techniques (e.g., rotations, scaling) to increase robustness .
Model Selection: We explored various models, including Convolutional Neural Networks (CNNs) due to
their effectiveness in image recognition tasks.Experimented with traditional machine learning models (e.g.,
SVM, KNN) for baseline comparison.
Model Architecture: Designed a CNN model with layers like Convolution, Pooling, and Fully Connected
layers for optimal feature extraction and classification.
Training & Optimization:
• Used cross-entropy loss and Adam optimizer.
• Tuned hyperparameters, including learning rate and batch size, for optimal results.
• Applied dropout layers to prevent overfitting.
SUMMARY
Project Overview: This project successfully developed a model capable of recognizing handwritten digits
with high accuracy.
Achievements: The model achieved over 98% accuracy on the test set, showing its effectiveness in real-
world applications.
Significance: Demonstrates the power of CNNs in image recognition tasks and the potential for deploying
similar models in applications like postal code recognition, bank check processing, and digitization of forms.
Key Takeaway: Leveraging CNNs, data preprocessing, and rigorous testing enabled us to create a reliable
solution for handwritten digit recognition.
FUTURE SCOPE
Improving Model Accuracy: Exploring deeper neural network architectures or ensemble methods to further
improve accuracy.
Real-Time Deployment: Implementing the model in a real-time system, potentially in mobile or web
applications for practical usage.
Multi-Language Handwriting Recognition: Expanding the model to recognize different alphabets or
characters, catering to multi-lingual document processing.
Cross-Domain Applications: Adapting similar models for other pattern recognition tasks, such as object
recognition, facial recognition, or anomaly detection.
ACKNOWLEDGMENTS
Mentor & Guide: Special thanks to Prof. Prabhjot Kaur, whose guidance and expertise were
invaluable throughout the project.
Team Members: Gratitude to each team member for their dedication and collaboration in making this
project possible.
Thank You
