Stuart Myles
Director of Information Management at the Associated Press
What is Image Recognition?
• Technology to recognize people, places, things or emotions in an image
• Available as APIs, as well as open source software
• Image recognition involves building a model to identify a set of topics
• Topics can be anything you want - baseball, happy faces, drug use, war…
• Requires lots of example images, so the software works out what patterns to look for
• Consumers of image recognition services are often stock agencies
• Off the shelf models are therefore available for concepts like
• Graphic/NSFW
• Celebrities
• Emotion
• General keywords – wedding, food
• Many commercial companies also offer to create custom models – for a higher fee
@smyles
Improve Search and Auto-Publishing
• Improve search experience
• In AP portals and in customer CMSes
• Keywords to match more queries and so surface more content
• Filters to narrow results to more relevant content
• Simplify auto publishing
• AP customers have fewer – or even no – editors to manually review content
before publishing
• Filters let customers fine-tune saved searches
• Eliminate customer need to manually identify graphic content
@smyles
AP Images Metadata
• AP handles 3,000 – 4,000 images a day
• Digitize 700 – 800 photos a month
• AP Images has about 34 million photos
• AP already applies metadata to images
• Manually by photographers and editors
• Mapped from third party feeds
• Automatically based on photo text – such as caption – via AP’s tagging service
• We manually keyword some archive images
@smyles
Early Days: First Half of 2017
• Early in 2017, we evaluated leading vendors
• None offered custom tagging
• Disappointing results
• Too many keywords that do not apply to image
• Inaccurate keywords scored with high confidence
• Some were strong for stock images, not so good for news
• Others were too generic
@smyles
High confidence:
Sunglasses
Woman
Man
Low confidence:
Finger
Hand
And notice: watch
High confidence:
Bat
Batter
Baseball
Softball
And notice: watch
Technology Evolves: Second Half of 2017
• We evaluated open source software
• Future option, but the software isn’t mature yet
• Later in 2017, vendors upgraded their offerings
• Most added custom tagging
• Working with business and sales, we designed a new image taxonomy
• Sports actions, NSFW filters, emotions, and image attributes
• Complements the existing news taxonomy we apply to text content
@smyles
A Hybrid Approach: Out of the Box + Custom
• Use out-of-the-box tagging for most concepts
• Train custom tagger for any concepts not covered (or covered well) by OOB tagger
• Find example images for each concept e.g. “tackle” - anywhere from 500 to 5,000 examples per concept
• Test the tagger to make sure it is accurate
• Feed the tagger more examples where it underperforms
• Proof image tagging in Production
• High confidence tags accepted as-is
• Ignore low confidence tags
• Medium confidence tags reviewed by Editorial
@smyles
Train and Test
Management
• We assembled
training sets that we
shared with the
partner
• And we held back test
images
• Testing for accuracy
• Precision
• Recall
Things We Learnt
• Assembling test and train sets is arduous
• But also where most of the value lies
• Some concepts are difficult to distinguish
• Dawn / Dusk, Happy / Jubilant
• Perceived concepts are different than text subjects
• May require some reorganization of our taxonomy and how we represent it
@smyles
Thank you!
Questions?
@smyles

Image Tagging at the Associated Press

  • 1.
    Stuart Myles Director ofInformation Management at the Associated Press
  • 2.
    What is ImageRecognition? • Technology to recognize people, places, things or emotions in an image • Available as APIs, as well as open source software • Image recognition involves building a model to identify a set of topics • Topics can be anything you want - baseball, happy faces, drug use, war… • Requires lots of example images, so the software works out what patterns to look for • Consumers of image recognition services are often stock agencies • Off the shelf models are therefore available for concepts like • Graphic/NSFW • Celebrities • Emotion • General keywords – wedding, food • Many commercial companies also offer to create custom models – for a higher fee @smyles
  • 3.
    Improve Search andAuto-Publishing • Improve search experience • In AP portals and in customer CMSes • Keywords to match more queries and so surface more content • Filters to narrow results to more relevant content • Simplify auto publishing • AP customers have fewer – or even no – editors to manually review content before publishing • Filters let customers fine-tune saved searches • Eliminate customer need to manually identify graphic content @smyles
  • 4.
    AP Images Metadata •AP handles 3,000 – 4,000 images a day • Digitize 700 – 800 photos a month • AP Images has about 34 million photos • AP already applies metadata to images • Manually by photographers and editors • Mapped from third party feeds • Automatically based on photo text – such as caption – via AP’s tagging service • We manually keyword some archive images @smyles
  • 5.
    Early Days: FirstHalf of 2017 • Early in 2017, we evaluated leading vendors • None offered custom tagging • Disappointing results • Too many keywords that do not apply to image • Inaccurate keywords scored with high confidence • Some were strong for stock images, not so good for news • Others were too generic @smyles
  • 6.
  • 7.
  • 8.
    Technology Evolves: SecondHalf of 2017 • We evaluated open source software • Future option, but the software isn’t mature yet • Later in 2017, vendors upgraded their offerings • Most added custom tagging • Working with business and sales, we designed a new image taxonomy • Sports actions, NSFW filters, emotions, and image attributes • Complements the existing news taxonomy we apply to text content @smyles
  • 9.
    A Hybrid Approach:Out of the Box + Custom • Use out-of-the-box tagging for most concepts • Train custom tagger for any concepts not covered (or covered well) by OOB tagger • Find example images for each concept e.g. “tackle” - anywhere from 500 to 5,000 examples per concept • Test the tagger to make sure it is accurate • Feed the tagger more examples where it underperforms • Proof image tagging in Production • High confidence tags accepted as-is • Ignore low confidence tags • Medium confidence tags reviewed by Editorial @smyles
  • 10.
    Train and Test Management •We assembled training sets that we shared with the partner • And we held back test images • Testing for accuracy • Precision • Recall
  • 12.
    Things We Learnt •Assembling test and train sets is arduous • But also where most of the value lies • Some concepts are difficult to distinguish • Dawn / Dusk, Happy / Jubilant • Perceived concepts are different than text subjects • May require some reorganization of our taxonomy and how we represent it @smyles
  • 13.