Joaquim Rocha
Published on

One more step in OCR with OCRFeeder 0.7

Authors
  • avatar
    Name
    Joaquim Rocha
    Twitter
  • Principal Software Engineering Manager at Microsoft

I have been hacking on some new and cool features on OCRFeeder for a while and now it is time to show them to the world in a new release.

These features I’m talking about fall mainly in 2 areas: improving the a11y of the UI and improving the recognition of documents.

A11y Improvement

The improvement of the a11y has the typical UI changes to include mnemonics, missing labels and relations, but also other approaches that have more to do with UX like using a progress dialog to inform users that time-taking operations are being carried. This means that now, the PDF importation and OCR won’t block the UI. Other changes in this category were the navigation through the content boxes (before, these could only be selected by clicking on them), the selection of all boxes and the deletion of selected boxes.

The following screenshot shows the box editor area of OCRFeeder with its mnemonics highlighted:

[Box edition area][1]
Box edition area

Recognition Improvements

Sometimes, text columns are so close to each other that they end up being recognized as a single paragraph, so I added a post-detection method to solve this issue. This feature is optional and can be toggled from the Preferences dialog.

Here’s an example of the difference it makes:

[Before columns' detection improvements][2]
Before columns' detection improvements
[After columns' detection improvements][3]
After columns' detection improvements

Scanned document images are usually skewed and this makes it more difficult for the contents to be successfully detected and “OCRed”. I decided to implement an algorithm to deskew these images. The algorithm uses the Hough transform to try to find lines in the image and their angles and, while it is a bit slow, it works well:

[Skewed image][4]
Skewed image
[Deskewed image][5]
Deskewed image

This action can be used in a loaded image but can also be configured to be automatically performed before the images are added. The Unpaper tool can now also be set to be clean images before adding them. This makes it much easier to successfully recognize images obtained from a scanner device.

Some fine tunning of the content boxes’ bounds was done by trying to shorten their margins, that is, lowering the distance between the boxes and their actual contents.

The font size recognition was also tweaked to solve the problem of having paragraphs with initials (you know, the huge starting characters) which were influencing the whole paragraphs’ font size.

To finish the recognition’s improvements, I have added an optional action to find and fix the text’s line breaks. Usually, OCR engines don’t consider “semantic line-breaks”, that is, OCR engines always insert a newline in the end of each line. Using some regular expressions, I try to find these “fake” line-breaks and recover the original flow of the text. Like some of the features mentioned above, this one can also be turned on/off from the Preferences dialog.

Here’s how the Preferences dialog looks like now:

Preferences_dialog

Preferences_dialog_recognition

To finish, images can now be dragged and dropped onto the pages’ area and the mouse wheel can be used to scroll horizontally combining it with the Shift key, thanks to Stefan Löffler, and of course, several bugs were corrected and code was improved.

As you see, this is a “rich” new version of OCRFeeder that keeps being the easiest way to use OCR in a desktop. You are welcome to file bugs in bugzilla or to send patches and features’ requests to its mailing list or approaching me if you’re in GUADEC.

Download: OCRFeeder 0.7 tarball on GNOME FTP