OCR of product label (barrel effect)

  • Last Post 20 August 2012
  • Topic Is Solved
cuvius posted this 15 August 2012

I would like to extract ingredients from the label of a beauty products, which are very often cylinder-shaped. It seems the "barrel effect" is giving the OCR engine some troubles. An example of one such photo is this:

Bottle with text

Currently I get the following result using English language and the textExtraction profile:

qua. axarnidopropytamine oxide, cNortwxidine
Quorate, isopropyl alcohol, oxidized potypropyte
polyethylene, hydroxy propyl mettiyloeW

(Removed some special characters which messed with the formatting)

Is there any way to use ABBYY OCR SDK for extracting the ingredients more precisely?

Order By: Standard | Newest | Votes
Dmitry Me posted this 15 August 2012

First of all, the image you present has very low resolution - the beginning of the word "oxidized" in "oxidized polyethylene" is barely readable after the original image is zoomed, so information is just lost and is beyond recovery. To address this you should get photos with several times higher resolution. That's the number one thing you could do to improve OCR quality.

Second, our OCR engine has preprocessing to address this kind of distortion but it is targeted for double-page spreads and typically deals with less distortion. The scenario you're dealing with - when the can has small radius of curvature - can be too much for that preprocessing technique.

cuvius posted this 15 August 2012

The original image has much better resolution (2592x1936), I just scaled it down to fit the question. Unfortunately that doesn't improve the results. Thanks for the swift reply, though, I realize this is a special use case.

sam posted this 18 August 2012

Are you working for Java language?

Dmitry Me posted this 20 August 2012

@sam: It shouldn't make any difference what language the OP uses - OCR SDK Service has a REST interface that is language and platform agnostic.