I need to read the digits of images like the one below:

Water meter

From what I understand, this will involve:

  1. Registering and segmenting the image into each individual digit
  2. Hand labeling a bunch of these individual digit images
  3. Training a classifier based on that hand labeled data
  4. Putting it all together

Am I on the right track? Am I correct in thinking there is no way to read the whole 7 digit integer at once? (Don't care about the decimal.)

If so, can anyone recommend resources on the best way to approach 1?

And what would be the best ML technique for 3?

Many thanks!

  • $\begingroup$ For 3 you can create your own haars cascades. You can find some examples in OpenCV $\endgroup$ – peter May 7 '18 at 14:06
  • $\begingroup$ Disclaimer: I'm not an expect in this field $\endgroup$ – peter May 7 '18 at 14:08
  • $\begingroup$ If you don't want to hand label individual digit images, you could employ unsupervised clustering technique which groups similar images together first, then you can just label the resulting clusters which should be much less manual work. $\endgroup$ – Billiska May 7 '18 at 14:44
  • 2
    $\begingroup$ What do you mean by "images like below"? Is it exactly one of images that you need transforming or only similar? What is your setting: is it always at that angle? Could it be axis aligned (that saves some processing)? Are this always the same digits? (same font, same device etc.)? Are light conditions always like this? Any technique depends on your setting, but it is not clear what is your use case and how much you can adjust on image gathering process. Number 3 depends on your setting or require full book to answer... $\endgroup$ – Evil May 7 '18 at 16:13
  • $\begingroup$ Sounds like you're on a reasonable track. I recommend reading about techniques for OCR; Tesseract and Abbyy; convolutional neural networks; methods for solving MNIST; image registration; spatial transformer networks; HSV representation. $\endgroup$ – D.W. May 7 '18 at 16:19

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.