How to preserve document structure in tesseract

Question

I am using tesseract ocr to extract text from an image. Preserving the structure of the document is very important to me. Currently tesseract does not preserve the structure

Answer 1

For multicolumn documents in which one wants to preserve a single column of continuous text (e.g., read column 1, then column 2) or documents with photos (e.g., newspaper articles) it's probably worth looking at adjusting the page segmentation method. The default page segmentation method with tesseract is to only do "Automatic page segmentation" but NOT "Orientation and script detection (OSD)."

Putting the psm setting to 1 tells tesseract to use "Automatic page segmentation with OSD." This allows tesseract to recognize a multicolumn document (rather than treating the page as a single block of text) and helps tesseract avoid trying to OCR non-text blocks like photographs.

For more on page segmentation methods, see: https://github.com/tesseract-ocr/tesseract/wiki/ImproveQuality

Here is a sample of the command line syntax to adjust the page segmentation method

tesseract imagename outputbase [-l lang] [-psm pagesegmode] [configfile...]

For more on the syntax, see: https://github.com/tesseract-ocr/tesseract/wiki

Answer 2

Tesseract code compresses spaces in output. You will need to change the code to preserve them. See Tesseract - ambiguity in space and tab post.

Answer 3

Adding --psm 6 option works in my case (command line)

Answer 4

The only reliable way would be enabling hOCR output and parsing it. It will contain positions of each word on the page in pixels, as in the original image.

You can do it by specifying tessedit_create_hocr 1 in Tesseract's config file, or in whatever API you use.

hOCR is a subset of HTML, and what Tesseract generates isn't always a valid XML, so you can either use an HTML parser or write your own, but you can't use reliably an XML parser.

Answer 5

Newer versions of tesseract (3.04) have an option called preserve_interword_spaces which should do what you want.

Note that the number of spaces tesseract detects between words may not always be the same between similar lines. So words that are left-aligned with a run of spaces preceding them (as in your example) may not be output this way -- the preserve_interword_spaces option does not attempt to do anything fancy, it merely preserves the spaces extraction found. By default tesseract collapses runs of spaces into one.

Details on this option are here.