I'm working on machine learning project that requires I collect hand-writing data of alphabets from an uncommon language (I'm 100% there is no available data out there). My question is how should I collect this data?

The approach I'm currently thinking of taking is asking about 20 people to write the letters (about 10 times for each letter) on a piece of paper and scanning that sheet, to have it in digital form. But what should I do next? Should I just vectorize the RGB value of each pixel? Should I use an entirely different methodology instead?

What are the textbook approaches and best practices?

  • what digital format is the scan? that matters for what the end solution(s) will be. – albert Mar 4 '16 at 3:38
  • @albert I might be able to make it go to jpeg. But usually it goes to pdf. But if goes to pdf I can just screen shot each page, giving me a png, which I can retain at that format or simply convert it to jpeg. No doubt pdf format is bad, but is there any difference in terms of convenience among the different image formats? – Ragnar Mar 4 '16 at 3:46
  • png is lossless, so that is your best bet imho – albert Mar 4 '16 at 14:12
  • You might want to take a look to see if the zooniverse software used for Ancient Lives is available. – Joe Mar 7 '16 at 17:05

The first step is splitting the image into character arrays. To do that, check out the answers in this question: Separate image of text into component character images. In particular, the ImageMagick answer from 2015.

(If you can determine how the input is given, then collect the characters as separate images.)

To convert the image into a 2D array, you can use Python with the Image and Numpy libraries- see for example

Your Answer

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

Not the answer you're looking for? Browse other questions tagged or ask your own question.