I just started on an open source and open weight supervised learning model to recognize japanese kanji characters drawn on the screen.

I have a working prototype written in Julia which is a very simple neural network. The input is in vector format so traditional convolutional neural networks don’t work out of the box but I swapped the convolution layer with a path simplification algorithm and it worked extremely well. Like 20 samples per character (from a set of only 5 hiragana during prototype phase) was enough to get 100% accuracy in a test collection of 5 samples per character after only 30 iterations of training.

I plan an working with free and open data, which I don‘t think exists for japanese kanji characters (at least not in vector format; KanjiVG only has one sample per character and I need dozens) so I also build a crowdsourcing web site to collect data from random people on the internet.

I am planning to run some more experiments with my prototype model before I release the crowdsourcing web page to an actual server though.

Model prototype: https://github.com/runarberg/kantoku-prototype

Crowdsource app: https://github.com/runarberg/kantoku-collector