Best resources for recognizing handwritten text?


(bhaarat sharma) #1

I came across https://www.hyperscience.com/forms where they demonstrate the ability to recognize handwritten text. I am interested in learning from folks in the deep learning field who know more than I do as to what techniques/resources are out there for me to train a network to recognize handwritten text?

I know there is IAM database http://www.fki.inf.unibe.ch/databases/iam-handwriting-database but beyond that I’m not too sure. What lessons from fast.ai courses would be applicable for this?


(Jeremy Howard (Admin)) #2

You’ll need to be familiar with pretty much all the material covered in parts 1 & 2 of the course to build such a system, since it incorporates language modeling, segmentation, object localization, classification, etc. You can get some ideas from the ICDAR handwriting competition entries: http://u-pat.org/ICDAR2017/program_competitions.php


(Sahil KARKHANIS) #3

Hello Jeremy, I am also very much interested in doing this handwriting recognition project but have few questions regarding this. What if I have an image only of one or two words in it, would I need a object localization for that? Also I am a bit confused about how to move forward with a training dataset, what dataset should I use for this ?

Because from what i have at the moment is a dataset which has all the letters Caps and small and the digits (0-9) in different folders. What my concern is, while training on this kind of dataset it would train on images only of a particular alphabet or a number but when it receives an image of a complete word, how would it go forward with this? Do we need a sliding window or something in this case?


(hari rajeev) #4

have a look at this CRNN implementation


(Alexandre Blanc) #5

Hi,

I might be a bit late but i think you’ll like it.
http://openaccess.thecvf.com/content_ECCV_2018/papers/Curtis_Wigington_Start_Follow_Read_ECCV_2018_paper.pdf

Hope it would help.