Handwriting-Recognizer

See this project here: https://handwriting-reader.herokuapp.com/

• Description: The purpose of this web-app is to recognize a number from 0-9. Users can draw the number in an input box on the website or upload a image of their number. Users can also help the network learn by giving feedback on correctness of our network’s guesses.
• Project Hierarchy:
  • Data: -Contains files downloaded from MNIST handwritten digit database
  • Static
    • script.js contains logic for user drawn input
    • style.css contains style information for templates
  • Templates
    • index.html contains the template for the site homepage
    • recognize.html contains the template for our networks guess. It also contains options for the user to help train our website by validating (or correcting) our guess.
    • read.html is a page we created to help debug our “blurred” user image. This page is not displayed in the working, submitted version.
    • appplication.py
  • Screenshots: screenshots for this document.
  • Flask application
    • recognize.py
      • Guesses number from user input as well as retraining data-set with user input
    • vector_helpers.py
      • Calculates vectors from user input and formats in array. Also attempts to make user-drawn input look similar to database training set format
    • bitmap_helpers.py
      • Contains helper functions that format user uploaded input
    • MNIST.py
      • Preprocessing for MNIST data set. Once this data is formatted into array, we return the array to train.py. This file is only run once and just kept for reference.
    • train.py
      • Uses keras to train the model. Training has already been done and this file does not run unless user explicitly enters command. File has been kept for reference and completeness sake.
    • modely.h5
      • Contains model’s trained data. This file is saved on the server and continually updates as users add training.
• Usage: go to https://handwriting-reader.herokuapp.com/
• On the website, users can either draw in the white box or upload an image to be recognized. Once uploaded, the user hits “Go” and is redirected to a recognition page.
• This page displays our networks guess. We also give the user the option to help our network learn. Simply type in the number that the prediction should be.

Credits: The software depends on the Keras library for Python3, which uses the Tensorflow library to create, train, and run the neural network. The training data was taken from the free MNIST database hosted here http://yann.lecun.com/exdb/mnist/. This project was inspired by the Detexify web app here http://detexify.kirelabs.org/classify.html.