A network for digit recognition?

Question

Text recognising

Is it possible to recognize numbers from such a captcha by a neural network?

Answer 1

of course, but it will probably be expensive with a purely neural network.
First you need to clear the image of garbage, obviously by first removing all black dots, replacing them with white ones, then white ones, replacing them with black ones, and then removing black areas into which you can fit a circle larger than a certain radius , after that you delete the remaining objects whose area is below a certain threshold, .. after which the numbers and long stripes remain (and maybe most of them will go away)
you look at the rotation angle of the numbers, finding a rectangle above a certain size in which the objects are inscribed, the angle of inclination of the long the sides will be the desired one, the contents of the rectangle are rotated and scaled to a constant size and further processed (the numbers will not be rotated but with a random offset)
further, it is already possible to select the numbers, according to their minimum and maximum width, looking for the minimum number of black dots on the vertical running scanning line.
it won’t be difficult to train a neural network to determine already processed numbers, and by the way, a neural network won’t even be needed here if you have a rectangle with an inscribed number, divide it into areas, look for the presence of black or white dots in them, which is more, and already compare them with in advance prepared list of options