This is a demo video. Please watch it before you try out the

demo site


The numbers 0 to 9 were tested to see the model can recognize the real handwritting.
It seems like it cannot recognize well when it comes to number 6 and 8. The reason might be its poor weights about judging number 6 and 8, due to overfitting. (The implementation is lack of regularization method.)
The accuracy used in this demo was around 0.8.
The weights were driven by the model that i made, not from keras.
If the weights were driven from the keras, then it might have shown better performance, however, the purpose of demo was not showing the performance of keras, but the algorithm i implemented and trained.
It can be improved by adding regularization.