I am trying to separate scanned pages of a 3 column book into paragraphs. On the pages there can be images located in an arbitrary location, occupying part of one, two or all three of the columns.
First, I started with regular 3 column grayscale pages without images and tried to locate paragraphs. For that I separate page into 3 columns:
- Calculating average of color (grayscale) along page width
- Threshold application to get equal almost whites to whites
- Peak detection
- Cutting page based on peaks
As it comes to paragraphs I have done a similar process in vertical direction on single columns. Which gives more or less satisfactory results.
All of this fails for any page with images. Many of my searches on the internet lead to OCR solutions, which 1. Do not address my problem directly 2. Are trained models for Latin alphabet based languages, which is not my case.
Are there any standard approaches to paragraph extraction from pages with images?