So the British Library has put together a dataset for image segmentation and OCR of Arabic manuscripts. I was wondering if anyone would be interested in working on this together. I worked on the project last year and am spending some time working on it this year.
It would be super cool if we could put together a solution using the fastai library to implement some deep learning approaches to these tasks and see how we perform.
If anyone is interested in working together, please give me a shout!!!
This is super cool. I’ve been following @tkasagi’s work on old Japanese manuscripts and I’ve wanted to do something similar with old Italian manuscripts that a professor friend has, but don’t know yet what’s feasible.
I honestly don’t know yet how much I can contribute, in terms of both time and of expertise, but I would like to contribute how I can!
I’ve put together a github repo with a few notebooks exploring the data and trying to train a basic model: https://github.com/oasis789/RASM2019.
The competition is split into three tasks - the first is page layout segmentation to identify marginalia, text and graphics. The second is text line segmentation and the third is the actual OCR.
I am going to try and focus on the first two tasks and build an image segmentation model using what we learnt from the lectures. I’m just running into a few issues because the images are really large and need to split into tiles for training and prediction. I’m not sure what the best strategy is for going about this and doing image segmentation with really large images.
Does anyone have
experience with image segmentation with large images?