MNIST real world digit recognition problem

Hi am a fellow hobbyist/project person & 6 months ago I undertook a project that required my mnist CNN to recognise real world handwritten numbers but it performed pretty bad so i spent 1 to 2 weeks of my time in the project trying to figure out a solution I searched on the internet & found nothing but eventually I came up with an idea then CNN recognition improved considerably but was still just not good enough because it was literally still failing on clearly written numbers like 7 so months later I thought to myself this CNN is better but could really be even better then the idea sparked in my head as to what’s causing my CNN to not recognise these digits easily then i confirmed the problem & implemented the solution now I literally have a CNN that performs 90%+ accuracy on real world data but to be honest im being too humble & im underselling my CNN it actually performs 99%+ on real world data

When I say real world data I mean numbers written on a blank page or a page that has blank ink but the hand written numbers are written in red or blue but it can be any other colour than black, if u write in black then it has to be on a completely blank page because in that way my binariser will be able to produce an image ready for processing ( it’s a long story but this is kinda off ramping so I won’t dive in it too much)

If u want the solution you have to pay for consultation simply comment in the comment section & we can get my solution tailored to your custom project so we can get your project back on track 💪

Видео MNIST real world digit recognition problem канала Tik Insta