(This documentation is still largely work in progress, use with caution)
The original paper refers to: Stroke Width Transform, Boris Epshtein, Yonathan Wexler, and Eyal Ofek 2010.
It is a long story, as always, please read their paper. SWT tries to capture the only text effective features and using geometric signature of text to filter out non-text areas. As a result, SWT gives you reliable text regions that is language neutral. Try it yourself:
./swtdetect <Your Image contains Text> | ./swtdraw.rb <Your Image contains Text> output.png
Checkout output.png, luckily, the text area is labeled.
SWT is quite fast. The SWT without scale-invariant support (multi-scale) can run on a 640x480 photo for well under 50 milliseconds on my laptop. By extending SWT to multi-scale, the accuracy increased by about 10% with about 2~4 times longer running time.
ccv’s SWT implementation performs on ICDAR 2003 dataset achieved similar performance with what Epshtein et al. reported in their paper, namely, with the old measure method described in ICDAR 2003 contest, ccv’s implementation was able to achieve precision rate at 66% and recall rate at 59% (numbers reported in the paper are precision rate 73% and recall rate at 60%).
However, these results are quite out-dated, and by using ICDAR 2011 dataset, more meaningful comparison is possible.
With ccv’s scale-invariant SWT implementation, and do parameter search on ICDAR 2011’s training dataset, I was able to achieve:
precision: 59% recall: 61% harmonic mean: 60%
Which would rank around 2nd to 3rd place in the chart. Please note that other methods in comparison are language specific, thus, were trained with additional character shape information using SVM or Adaboost where as SWT is language neutral and doesn’t use any language specific features.