I have a big complaint about function TextRecognize. Alexey Popkov have verified here. Actually it can do it as the attribute "BoundingBox" if the function is normally.
HighlightImage[ColorNegate[Binarize[img1]],
TextRecognize[img1, "Word", "BoundingBox"]]

But as you see,it will give a very bad result.Since the MMA have the function of neural-networks,so I want to expect a neural-networks method to find all number and the position of the number. I have some samples image here for test
img1=Uncompress[FromCharacterCode[
Flatten[ImageData[Import["https://i.stack.imgur.com/keYUf.png"],"Byte"]]]]
img2 = Uncompress[FromCharacterCode[
Flatten[ImageData[Import["https://i.stack.imgur.com/U86l9.png"],"Byte"]]]]
img3=Uncompress[FromCharacterCode[
Flatten[ImageData[Import["https://i.stack.imgur.com/V76u5.png"],"Byte"]]]]
Could anyone can give some advice?
ContourDetect[img1]to detect the position of the numbers and characters, and then crop out the numbers and characters and feed them into a classifier. – xslittlegrass Sep 22 '17 at 17:33