1

I'm looking to recognise a set of handwritten numbers, starting from 0 and increasing by 0.5 for every number i.e. 0, 0.5, 1, 1.5, 2, 2.5, 3, 3.5 ... 10. Tried searching online but to no avail; MNIST does not seem fully useful either given it only deals with whole numbers.

Right now the plan is to augment MNIST's dataset with a few thousand images of the n.5 numbers with a few people's handwriting in various styles. The trained model would be used to recognise the same numbers written by other people. My concern is accuracy could be low given the uniqueness of every person's handwriting. Are there any alternatives which is more efficient or possibly produce a high accuracy rate? Thanks a bunch.

dasanibottle
  • 9
  • 2
  • If you can recognize any digits and a decimal separator, what prevents you from recognizing any arbitrary number? – Richard-Degenne Sep 29 '17 at 10:07
  • @RichouHunter I presume by recognizing any digit and a decimal separator, you mean recognizing the digits and decimal separator as discrete objects. That sounds like a viable option, if we can use two adjacent boxes to put the numbers in, the left box representing the ones and the right representing the tenths. We can do without recognizing the decimal separator and still make use of MNIST's database. One shortfall is that '10' can no longer be identified, but we can always have an additional box if needed. – dasanibottle Oct 02 '17 at 01:25

0 Answers0