Web Application Repository (Frontend + ML Integration):
GitHub – Web App
Live Web Demo (Digit Drawing & Prediction):
Live Demo
Project Webpage Repository:
GitHub – Project Webpage
Project Webpage (Overview, Results & Report):
Project Webpage
This project was done as part of CSL2050: Pattern Recognition and Machine Learning at
Indian Institute of Technology, Jodhpur,
Spring Semester, Academic Year 2024–25.
- Project Description
- Methods and Performance
- Results
- Video Walkthrough
- How to Run the Project
- Contributors
This project focuses on building a Handwritten Digit Recognition System using classical machine learning techniques. The objective is to design, implement, and compare different classification algorithms for recognizing handwritten digits with high accuracy.
The dataset used is the MNIST (Modified National Institute of Standards and Technology) dataset, consisting of:
- 60,000 training images
- 10,000 test images
- Image size: 28×28 pixels
- Labels: digits 0–9
MNIST is a standard benchmark dataset widely used in computer vision and pattern recognition tasks.
We implemented and evaluated six machine learning algorithms using the same training and testing splits for a fair comparison.
A non-parametric algorithm that classifies data points based on the majority class of their nearest neighbors using distance metrics such as Euclidean distance. While simple, KNN can be computationally expensive for large datasets.
SVM finds the optimal separating hyperplane between classes and can handle non-linear boundaries using kernel functions. It is particularly effective for high-dimensional data and achieved the best performance in this project.
An ensemble learning method that combines multiple decision trees trained on random subsets of data to improve generalization and reduce overfitting.
A linear classification model extended to multi-class classification using a one-vs-rest strategy. It estimates class probabilities and selects the most likely class.
A tree-based model that recursively splits data based on feature values. While interpretable, decision trees are prone to overfitting if not carefully regularized.
A probabilistic classifier based on Bayes’ theorem with the assumption of feature independence. Despite its simplicity, it performs competitively for digit classification.
| Algorithm | Accuracy (%) |
|---|---|
| K-Nearest Neighbors (KNN) | 86.53 |
| Support Vector Machine (SVM) | 95.44 |
| Random Forest | 92.25 |
| Logistic Regression | 92.57 |
| Decision Tree | 86.52 |
| Naive Bayes | 88.15 |
SVM achieved the highest accuracy, followed by Logistic Regression and Random Forest.
A detailed walkthrough explaining the project architecture, models, and web interface:
(Please set video quality to at least 480p)
You can try the project directly using the web interface:
🌐 Live Demo:
👉 https://your-demo-project.vercel.app
- Draw a digit using the on-screen marker
- The trained models predict the digit in real time
| S. No. | Team Member | Contribution |
|---|---|---|
| 1 | Nitin Verma | SVM Implementation |
| 2 | Nitesh | Logistic Regression Model |
| 3 | Shruti Sunil Vibhute | Random Forest |
| 4 | Saurabh Kumar | Naive Bayes Model |
| 5 | Vanshita Jeenwal | Decision Tree |
| 6 | Mallam Vishnu Priya | KNN |