Mausr - Marek's Unicode Symbols Recognizer
Mausr is a library and web interface that performs unicode symbols recognition from hand drawn input using a simple neural network. The symbol recognition library is written from scratch without any ML library and it implements a back-propagation learning algorithm as well as basic non-linear optimization algorithms such as gradient descent and iRprop.
Source code is available on GitHub: NightElfik/Mausr
Summary#
Unfortunately, I have never got time to write a detailed article about this project and the original website is now down. If you are interested, see the full source code at GitHub.
Highlights of the project are:
- Uses a neural network for image recognition, used for unicode symbols search.
- Implements back-propagation, gradient descent, and iRprop algorithms.
- Web interface for interactive training and symbol search using hand-drawn inputs
- Ability to collect new training samples from users.
- Statistics about trained model.
Recognition results. 
Interface for training neural nets. Real-time feedback of the training is implemented using the SignalR library. 
Results of training a neural network. 
Test of optimization functions. 
Drawings of a snowflake from many users.
This post is licensed under CC BY 4.0 by the author.