Many papers mention that text regions give rise to high eigenvalues calculated from the greylevel pixel values after the image has been divided into blocks of mxm matrices. Also that eigenvalues are a measure of the ‘roughness’ of the texture of the image.
How is that related to getting text? Text areas are generally of two colors, background and foreground with the letter-strokes of uniform color. Where is this roughness – there could be many other features that would be more rough and trigger high eigenvalues. Could someone point out where to get the math that connects these things?
EDITS:
A few papers included that mention eigenvalues in the context of text detection in natural scenes.
-
A framework towards realtime detection and tracking of text uses the Eigentransform on greyscale image.
-
An eigenvalue-based approach to text detection in video mentions calculating eigenvalues from the covariance matrix of gradient image.
Just an orientation, so you can start reading and eventually target your next question better:
You are talking about Principal Component Analysis
Here you have an example application:
HTH to get you started.