MNIST Handwritten Digit Classifier Web app

Overview

This project implements three different methods for classifying handwritten digits from the MNIST dataset: Deep Neural Network (DNN), Data Augmentation with DNN, and Data Augmentation with Convolutional Neural Network (CNN) in MNIST_Classification.ipynb. The goal is to explore various approaches to improve classification accuracy and evaluate their performance. The best-fitting (CNN) model is used for prediction: CNN_Augmented_100_model.h5 in the web app.

Live Demo

This interactive app enables users to upload an image of their choice, and it predicts the (handwritten) digit in the image. Deployed at Digits-In-Ink using Streamlit.

• Usage :
  • Please use a Desktop environment for the best experience. Either manually browse for image files by clicking on Browse files option or try with some real life images present below the File browser by dragging and dropping onto the Browse files option.
  • Then click on Predict button to get the predicted digit. You can also view the class prediction probabilities in the expander below.
  • If you are frequently getting incorrect results try following the steps given in the expander at the end.

NOTE: If the webapp link doesn't work properly, run the app on localhost by following the below steps.

• For running on localhost : Prepare the environment by installing all the requirements from requirements.txt ( * pandas and matplotlib are required only for the ipynb file so the app will work even w/o those libs). Then type the following command in the terminal : streamlit run app.py

Screenshots of app

Homepage

Locally browse images

Prediction along with class probabilities

Methods Implemented in ipynb

1. Deep Neural Network (DNN):

   • A basic feedforward neural network architecture.

2. Data Augmentation with DNN:

   • Utilizes data augmentation techniques such as rotation, translation, shearing and scaling to increase the training dataset's size artificially.

3. Data Augmentation with Convolutional Neural Network (CNN):

   • Incorporates convolutional layers for feature extraction and spatial hierarchies, combined with data augmentation techniques.

Dataset

The MNIST dataset is a classic benchmark dataset consisting of 28x28 pixel grayscale images of handwritten digits (0-9). Each image is labeled with the corresponding digit it represents. More info can be found at the MNIST homepage.

The dataset images can be downloaded and split into train and test sets simultaneously with the code :

(x_train, y_train),(x_test, y_test) = tensorflow.keras.datasets.mnist.load_data()

Steps to run the MNIST_Classification.ipynb

• Upload the MNIST_Classification.ipynb to Google Colab and run the notebook (As Colab was used originally for training purposes)
• To run the ipynb on a local machine, remove/ comment-out the code: from google.colab import drive; drive.mount('/content/drive') and set the paths to save and load the model accordingly.
• To test the model without training, use the .h5 model files in the models directory.

Results

The dataset was almost balanced (not biased).

Methods used	Accuracy	Loss
DNN	97.97%	0.1070
Data augmentation + DNN	98.72%	0.0365
Data augmentation + CNN	99.45%	0.0153

Contributions and suggestions are welcome! 😊