ALIPR - Image Recognition for Keywording
Image recognition - that is - the automated determination of the visual contents of an image parsed into text, is the holy grail of search. Back in July, we wrote about it (Image Recognition and the Future, 7/2/08) and several very interesting advances - and for you bleeding-edge readers out there, a behind the scenes look at how.
With the speed of Moore's Law applied, fast-forward to present day, with the here-to-fore unknown "Automated Linguistic Indexing of Pictures Real Time" - or ALIPR. (note it seems to be missing a "T" in the acronym.) Now, it launched in 2006, but it's just hit my radar this weekend, thanks to the good folks at TechRadar.com (Automatic image tagging gets smarter, 10/9/08), and LifeHacker (ALIPR Learns How to Auto-Tag Photos, 10/11/08), which also has other examples.
Here's one example of a fairly successful (but by no means exhaustive) result return, where I've checked the applicable keywords (click each to see them a bit larger):
Here are several others with less than stellar results. Of particular interest is what it DID return. Stop and think about how it came to choose the keywords it did - patterns, colors, and so forth:
As this technology gets more effective, there will be an inverse law applied to the technology from a keywording perspective. Just as we are on the lookout for an amazing automated piece of software that will generate keywords for us (like this one, albeit in it's admitted infancy) so that our images can be more easily found by those needing to find them, we will not need keywording (or atleast need it less than we do now) because we will have search results that return the images we need, even without the keywording.
To give it a spin yourself, visit them here.
Please post your comments by clicking the link below. If you've got questions, please pose them in our Photo Business Forum Flickr Group Discussion Threads.
1 comments:
John:
I had first brought ALIPR to the attention of my Controlled Vocabulary forum (http://www.ControlledVocabulary.com/forum.html) in the fall of 2006.
According to their website (then and now), ALIPR uses a vocabulary of 332 English words. I fed it an image of a young boy yelling (the same one used in the IPTC Core sample images) in November of 2006, and was asked to check boxes next to those terms that were correct. Out of the list of: landscape, building, desert, historical, rock, animal, wild life, indoor, bath, kitchen, people, cloth, female, face, and glamour; only "people" and "face" were correct. I repeated this test with the same image in the spring of 2008, with only one correct match as a result (people). One would think that over a years time the system would have "learned" and gotten better, but this does not seem to be the case.
Several other images I tested at the same time did fare a little better, but the best results still only yielded three correct matches per image.
If you take a look at the most recent 15 images (http://www.alipr.com/cgi-bin/alipr_recent.cgi?n=15) and take a look at the "guessing" done by ALIPR, you’ll see that this technology is still in the early stages of development.
Check out the UK startup Imense, http://www.imense.com/ for another take on how this technology is progressing.
David Riecks
--
Need Keywords for your database? Get the Controlled Vocabulary Solution
http://controlledvocabulary.com/products/
Post a Comment