If we could automatically pull any information from the document or even just know what type of document we’re looking at ahead of time, we could make our document processing even more efficient. Every second saved over the course of 26 million documents a year would bring huge savings for our process.
C.H. Robinson Engineering, “Document Detection Using Data Science“
The document classification model I put together for this effort was followed up by an experiment to see if we could determine whether a document had been signed. This went beyond the handwriting vs. printed text models you’ve probably seen elsewhere, because these documents would almost always have initials, scribbled notes, etc. but what we really wanted to know was if someone had actually signed for delivery.
Turned out to work well in practice.
To show that the model was able to differentiate handwriting vs. initials vs. signatures, I applied a transparency transform to the above image where the transparency was proportional to the model’s confidence it was looking at a signature – the more certain the model was looking at a signature at that position, the more opaque I made that part of the document.
You can see that the model was able to ignore pictures, printed text, etc. and even though it showed a bit of confusion with handwriting its final prediction was correct.
