[Tutorial, Long-term Update] Handwritten Chinese Character / Number Recognition (MAIX BIT)

Written in front: Hello everyone, I am Chengsen Dong (xddcore), ee is studying, and I have also tossed a lot about neural networks and machine learning in the last six months. (Algorithms are very good, they are all adjusted to High-level API). I currently want to port some of my demos to the k210 (of course Micropython). Yesterday I saw you have some problems with multi-digit recognition. Then let’s transplant the demo of multi-digit recognition of opencv + tf. In this post, I will also share how to train a model on a PC and deploy it to the k210. And some ideas for optimizing the model.

The following are the effects:

About dataset:

Regarding the digital identification data set using mnist, and then the Chinese character identification data set using CASIA-HWDB collected by the State Key Laboratory of Pattern Recognition, Institute of Automation, Chinese Academy of Sciences in 2010.


1. I will fuse these two data sets, and finally become a data set that can recognize both numbers and handwritten Chinese characters. Then throw it to the convolutional neural network for training.

2. Try to use the model conversion tool provided in Maix’s toolbox to finally convert the model to Kmodel. (In this process, it may take some time (the landlord’s model optimization ability is basically 0))

3. runing in maix bit.

1 Like

Hello xddcore,

Excellent approach!
I am very interesting in your results. Especially the transfer from Keras on PC to the K210 and the Maixpy code are important for me.

Best regards