Hacker News Reader

Comment by Trun_wal | Hacker News Reader

Hey, amazing blogs. Currently working on degraded scanned documents. Are there algorithms distinguishable for documents and natural images?

I am using open cv to process the documents, curious if I am missing out chunk of cv algorithms specially for scanned administrative documents (financial,personal documents)?

View on HN

I'm not sure what you mean by "algorithms distinguishable for documents and natural images" -- can you elaborate? OpenCV itself doesn't have builtin functionality to take documents and fit them to a pre-defined template, that tends to be part of a specific use-case/niche of computer vision for document processing. The general idea is to take a document a user has filled out and "fit" it to a blank template, where you know exactly where each field is. That way you can exact the information from the document.

"The general idea is to take a document a user has filled out and "fit" it to a blank template" - I agree point to point. However, I am struggling with templatization due to poor quality of the document images. To process those documents (denoise, super resolution, HE - etc. etc.), the OpenCV algorithms are not working good enough, requires a lot of tuning varying with each document.

So, I was wondering if those algorithms work better for natural images (buildings, people, things etc) than document images (text, graphics) and if so, there must exist algorithms to process such documents I am unaware of.

Reply on news.ycombinator.com