It is as you said, use the crop method to make new image from each of the characters, barbarize the cropped image, then draw that cropped image in the middle of a white area with Same of width/height pixs.
What I feeding the network is exact what it’s trained on. Such as trained it use 32x32 sample, then feed it with 32x32 pictures. So before feed, I must pre-fix Each Roi’s Picture. That’s all.
Wish CMSIS-NN can be used in future. Now it works with 20 characters only 1.3 FPS. Still can’t use in real application. Is it true if save it to the CPU as main.py, it can run fast ?
Some bug report: when I min the window and then max the window of IDE, all the sub window get back to the original size, not the custom size. and sometime the menu (such as File), when click it don’t display the down sub menu.
I think CMSIS-NN is big, If the IDE can deploy different FW based on users intent, maybe get the memory free.