OCR/Roadmap

Currently, all products are edited manually. This project is about automatic or semi-automatic detection of a number of things using OCR and Computer vision.

Current state

• OCR extraction of Ingredients using Tesseract 2 (production) and 3 (.net)
• Uses the French dictionary for all languages

-- /home/off-fr/cgi# grep get_ocr *
Ingredients.pm:use Image::OCR::Tesseract 'get_ocr';
Ingredients.pm: $text =  decode utf8=>get_ocr($image,undef,'fra');

• Has a small custom dictionary for French ( /usr/share/tesseract-ocr/tessdata/fra.user-words)
  • https://code.google.com/p/tesseract-ocr/wiki/FAQ#How_do_I_provide_my_own_dictionary

Short term goals

• Use the right standard dict for each language
• Integrate custom lists from Global Ingredients Taxonomy
  • Create a golden set
    • e.g. someproduct.jpg -> ingredients image
    • someproduct.golden -> ingredients text
    • then we create a script that runs the OCR through the images, compare with the golden text, and report some accuracy measures
    • Draft Script: https://lite6.framapad.org/p/OFF_OCR_Script
• Integrate custom lists from the live instances; language per language.
  • http://de.openfoodfacts.org/zutaten
  • http://uk.openfoodfacts.org/ingredients + http://us.openfoodfacts.org/ingredients
  • http:// fr.openfoodfacts.org/ingredients
  • USDA UNII list of ingredients (will also work for Open Beauty Facts)
• Process all images and make products searchable, even if not filled yet

Long-term goals

* Get dictionaries translations from Wikidata
* Investigate Ocropus for complex layout extractions
* Investigate Open CV for detection of patterns, logos…

Targets

• Logos of brands (Getting them from POD ?)
• Logos of Labels (standardized)
• Text (distorted - bottle case, diagonally - with low light, bright light)
• Standardized layouts (US Nutrition labels)
• Standardized text (quantities, EU Packaging codes)
• Barcodes (extraction in uploaded images)
  • Store in separate image for further reference
• Image orientation: check that the text is properly oriented to guess if the image is properly oriented.
• Deep Learning
  • Product photo on packaging - guess category based on product picture
  • Container: guess whether it's a bottle, cardboard…

Extracting areas is already great work: if we can extract logos or patterns, it will be faster for humans to double check and turn that into text.