Scaling factors comparison for Ocrad
5th of November 2007
Short story: 3
Longer short story: The best scaling factor for Ocrad is 3
I've been playing with Ocrad to pick out words from images that contain text. In this experiment I've taken screenshots of my Firefox and run those screenshot images through Ocrad to see what scaling factor parameter works best. Generally, the higher the scaling factor the better the results but also takes much longer time.
"GNU Ocrad is an OCR (Optical Character Recognition) program based on a feature extraction method."
What my little experiment entailed was to copy the text of three web pages as well as taking screenshots of the pages as PNG files. I also copied all PNG files to JPG files with 85% quality. Then I wrote a little program that spits out commands on the shell to scan these images, compare the results and take the time how long it takes to complete.
Here's what the results looked like for all 6 images:
Quite obvious conclusion from this is that the best result starts at scaling factor 3. The two poor performers at the bottom of the graph I can't explain. Two of the images were simply much harder to get right.
Nothing is obvious from this graph except the higher the scaling factor the longer it takes.
Combing the two graphs with a pinch of common sense leads you to a simple conclusion: the best scaling factor to use is 3.
If you want to try it yourself you need ChartDirector installed. Here are the experiment files which for the curious also contains the images I used in the experiment.