Distinguish text objects and arbitrary noisy patterns in degraded document images

I am looking for some *features* that could helping me to *distinguish between text object and arbitrary noisy patterns in degraded document images*.

This is an example of a document with some parts of noise, I have extracted manually 4 windows, three of them correspond to noise, and the last one to text.

- noisy image

- noisy area 1

- noisy area 2

- noisy area 3

- text area

Is there any measures or methods that could do this kind of detection ?

Thank you.