Hacker News Reader

Comment by lalaland1125 | Hacker News Reader

Why hasn't this been automated? Digit recognition technology is pretty good and it seems like the forms are pretty standardized.

View on HN

Is there a technology that would help ocr-ing standardized forms? I’ve got a couple thousand pages of historical train schedules that i would like to digitize (tabulated, printed data, including symbols/icons), but I’m not sure how to automatically recognize the structured data.

I don't know about out-of-the-box ready software. Depending on how much time you have you'd either train your own model with your own ground truth or you can use an open source model. Libraries and archives sometimes make theirs available (I know the Dutch National Archives do). You can train the model and/or HTR the data with software like Transcribus or Kraken (open source).

(disclaimer MS employee but nothing to do with this product)

You could check out the Azure Forms Recognizer. It's designed to do exactly this and make even training a custom model pretty easy. It's also pretty cheap (free up to 500 pages) so you can always do a quick experiment and see if it does what you need it to

https://docs.microsoft.com/en-us/azure/cognitive-services/fo...

I remember getting some reference software and methods from NIST maybe 20 years ago on handwriting recognition. Never put it into practice, but maybe they have some updated software available

I'm curious, what's your interest in the train schedules?

One day at an antique shop, I came across a book from ~1910 which had hundreds of pages of annual reports from railroads with many metrics we'd expect to see in the 10-K reports public companies file.

The book was published annually, but had much of its data in tables with grouped headers and cells, which could make automated OCR-ing with a good (useful) end result challenging.

I think it'd be interesting to map out the Railroad consolidation, track all their financial metrics over time, and do some level of forensic accounting to see if/which companies probably had funny business going on.

I’ve got European rail schedules, which gives an indication of historical travel patterns. Specifically looking into the development of overnight travel.

Probably because 99% accuracy isn't good enough.

What accuracy do you think humans have?

If I give you 1000 handwritten numbers, do you think you'll make less than 10 mistakes?

They will likely assign the same tasks to multiple people and use an ensemble of human consensus.

Run both OCR and human in parallel and double check anomalies

I would expect humans to make similar errors.

AIUI the classic procedure for tasks like this is/was:

Hide some graduate students. Tell each to transcribe, and to mark the difficult spots. Give each page to two students. Next, have someone else process the page pairs and resolve conflicts and the marked trouble spots.

As long as one of the students notices that a particular spot is difficult to read, the error is discovered and can be handled by someone who isn't numb from transcribing pages of simple numbers.

The forms are in open access, anyone can fill one form, so they surely did consider variations in accuracy (if not plain sabotage). They are probably submitting the same data to multiple people and the cross-checking the submissions. Also, the totals help to filter out submissions with mistakes. One can also add bound checking for individual values (misplaced decimal point, extra digits etc.). I would love to see how they do it, but they'll probably keep it confidential because trolls and other vandals.

This is exactly how zooniverse works. The first thing you do is go through a few known pieces of data to verify you know how to identify things correctly (more important for galaxy types, etc).

As you start labeling data everything has to reach a consensus. I'm not sure exactly how it works, but it does have multiple people verify each piece of data.

Haha 99% accuracy in weather data would be an amazing improvement. Noisy data and data holes are all we got to work with. Humidity sensors have a memory, storing moist/dry air inside the ceramic.

Checking the results of OCR is a lot faster than manually doing the OCR, no?

I've had a look at a few of those and I doubt you'd get 99% accuracy. A lot of the footnotes especially are very hard to decipher. They're also 50 years old, handwriting has changed over time. I'm not sure how well models are trained to read old handwriting, I know that I struggle with it.

For humans, much can be deciphered through context. That's much more complex to do with OCR.

The National Archives in the UK has a research project to try to read > 200 year old handwriting with OCR reliably. They use various tricks like training per-century or per author models which gives a big performance improvement. They also have ways to handle symbols that no longer exist (not in unicode). Give me a shout if anyone here wants to collaborate.

I work on the visualization and publishing side of this kind of data, but I do know our HTR works even better if we combine authors and centuries into one model, instead of separating. Did you try and compare results?

I doubt you would get 99% accuracy with humans, either.

Outliers would be easy to spot.

Maybe they should split in lines each figure, make a coordinates map of the entire figure or so, make a copy and applying a bulk search with a machine for crossing the map and annotate as many undoubtely identifiable numbers as possible. Then paint it in a different color easy to filter or hide it and remember its position.

And then add humans to focuse only in the remaining dificult cases and outliers armed with a reference sample chart. This way an human would need to focus its eyes in 20 characters/image (instead 60 or 100). They would accumulate more completed figures faster and obtain a bigger sense of reward. If the human can't recognize the number, could put a ? in the chain and move on.

Thus the machine would annotate for example 1_3, 4_67_, _8 and the human would write: 5?01 for the same line.

Just an idea, don't know if designing that is easy-peasy or really defiant but the number of possible values is limited in any case, so should be possible to train a machine to recognize some single characters or even entire numbers. Specially if written for the same people.

Reply on news.ycombinator.com