• Python
• Neural Networks
• HTML
• CSS
• JavaScript

This project was a website where users could learn and test themselves on Japanese hiragana knowledge. Before duolingo introduced character drawing, there didn't seem to be any existing tools for users to be able to draw characters, and have them marked based on stroke order.

I decided to implement a neural network that would decide which character it thought the user was drawing, also taking into account stroke order (the order in which each 'line' of the character is drawn) which is important for proper technique.

The program took the large image (drawn on a canvas object), and converted it to a 28x28 pixel image, where the RGB value of each pixel was taken to feed into the neural network. This was done after every stroke that the user inputted. This had a 90% accuracy rate of detecting when a character was correct.

In addition to the neural network, there was also a simple matching game, and learning sections for these characters.

I had never used any sort of neural network, and this was my first journey into web development - and therefore this project developed some core software engineering skills for me! As well as learning about all of the technical aspects, after this project I was inspired to continue learning Japanese. I have continued independent and academic learning of the language and really enjoy it!!

There were no appropriate datasets available for taking into account stroke order for characters that I could use to train the neural network on, and I therefore drew them all myself - around 50 of each character, which were then distorted/translated slightly to increase the number of each character that was used.

I would like to extend this project to also work with katakana and kanji characters (the other two japanese writing systems). This would have to be gradual development, and only for a subset of kanji to begin with - as there are a LOT of characters.

Link to Git:Here