Show HN: Neural network optical character recognition | Hacker News Reader

Show HN: Neural network optical character recognition

github.com

23 pointsmegalodon9 comments

Are the reported success metrics on the training or testing set? The website says its on the training set, which shouldn't be a valid metric of success since neural networks can easily overfit to their training data (one of their downsides if you aren't careful).

Having the output layer be an 8-bit character representation though is very clever, rather than a softmax layer with each node being the relative probability of a given character. That probably lowers the number of free parameters you have to train, which probably speeds up training and can help prevent overfitting. I'm interested in knowing what the true success rate is with this approach as it seems clever.

Btw, what's your loss function on the output layer?

Yes, the success metrics are measured on the training set.

I just ran a measurement with a separately generated testing set consisting of 52k characters which yielded a success rate of 98.53% on letters a-z. This turns out to be almost exactly the same as the rate presented on the repo page (98.52%). Will upload the result for your perusal as soon as possible.

You will have to browse the synaptic neural network library [1] for an answer to your second question.

[1]: https://github.com/cazala/synaptic

EDIT: A separate testing set is now generated in addition to the training set. I updated the success rates of the examples in the readme accordingly.

A training set like Google's Recaptcha data would be useful. Maybe Project Gutenberg, Wikipedia and other open source projects should start an open Recaptcha-like service to collect such data based on scanned documents/books/etc.

Google has released street view house number data. It's mostly digits, but it's a start.

600,000 digit images (labeled data):

http://ufldl.stanford.edu/housenumbers/

Won't an easier way will be to copy a large text from net, paste it into MS Word. Export each page as image. And then train it using that image and the known text.

It won't help that much, as every character/letter/digit will be of the same font face, e.g. Arial. Then it can only recognize scans with Arial font.

One needs a real world training set, a lot of variation, like house numbers and scanned books with labeled datasets.

I'm planning on adding a JSON config file to be able to tweak the network and add a lot more variation to the generated training set. Meanwhile, if you want to train it with several different fonts you can just add them to the 'fonts' array in captcha.js.

Shameless plug: Similar thing for GPU https://github.com/pannous/caffe-ocr

Reply on news.ycombinator.com