OCR Tips

From iDigBio
Revision as of 13:41, 2 October 2012 by Dpaul (Talk | contribs)

Jump to: navigation, search

FineReader tips

SG: Consider image size/color: 

What works best:

Size, Color/Grayscale, and Resolution:

Recommended Image Resolution: 300 dpi for typical texts (printed in fonts of size 10pt or larger), 400–600 dpi for texts printed in smaller fonts (9pt or smaller). For best OCR results vertical and horizontal resolutions must be the same. See User's Guide for additional information.

  • Fullsize, color JPEG images of herbarium specimens are ±7-15 MB in size and take about 2 minutes to process each.
  • Fullsize, grayscale JPEG images of full herbarium sheets are about 1 MB in size and take about 1 minute to process each.
  • JPEG images of primary collection label only are about 300-600 KB in size and take seconds to process each.

Pattern training: If using this tool, be sure to train the tool on an image that is the same resolution as the other images you wish to OCR.

What to look out for:

  • Setting the resolution too high (over 600 dpi) increases the recognition time. Increasing the resolution does not yield substantially improved recognition results. Setting an extremely low resolution (less than 150 dpi) adversely affects OCR quality. See User's Guide for additional information.

Recognition Server

ABBYY Effective Practices & Hints for users of Recognition Server - what works best in our experiences.

EH: start with fewer languages selected since each language adds to the time taken (potential to sort specimens geographically prior to OCR). We are currently processing our specimens from SW Asia and the Middle East with a large number from Turkey so we currently run ABBYY with Turkish and English selected.
EH: we select high quality rather than speed

PL: OCR quality can be enhanced when a large image is cropped - which also reduces page count.
PL: Images can be ingested from a shared folder, or scan station, or ftp/ftps, or API. Hotfolder ingestion can be further controlled by including an optional XML ticket. XML tickets control workflow, output, and allow metadata to be ingested along with the image to be processed.

- what to look out for (with examples).

EH: we find that running the whole image increases the page count, each image ends up as 4-6 pages. Not necessarily a problem but good to be aware of if page count is an issue.

Tesseract Tips

Tesseract Effective Practices & Hints.

What works best:

Resolution: x-height (pixel height of lowercase letter) between 20-40 pixels is ideal
Switching to grayscale, increasing contrast, and other image treatments can improve output at times

What to look out for:

Resolution: an x-height below 8-12 pixels will produce very poor OCR return
Using a black background for package labels (e.g. lichens, bryophyte) will create a black border that can significantly reduce OCR return
Form labels can interfere with OCR output
Faded labels or images with poor lighting can be problematic
Old font can be problematic. However, it is possible to train Tesseract for new fonts

Misc notes:

Will recognize vertical text
Image input can be tif, jpeg, or gif