Who among us has not gotten to a grocery checkout counter with a less
than common vegetable only to be confronted with a clerk who has no
idea what the item is, and, therefore, cannot figure out how to key
it into the system?
To help deal with this little frustration of life for the customer
and big problem for the retailer, IBM researcher
Jonathan Connell and
his colleagues in the Exploratory Computer Vision Group have come to
the rescue with VeggieVision, a high-tech version of the bar code
scanner that will recognize fruits and vegetables and automatically
enter the price into the checkout system. An added benefit could well
be shortened waits in supermarket checkout lines.
The project, which was on display at Expo 2000, had a false start
some seven years ago when the IBM cash register was a hot commodity.
It then lay fallow for several years before it was reincarnated in a
different form about a year ago. The first version consisted of an
ungainly black box that included
special lights to control the color of the vegetables as they posed
before a $1,000 camera. The visual information was fed for analysis
to some DSP cards that cost thousands of
dollars. With the advent of fast, low cost processors and cameras the
VeggieVision idea took on a new appeal and was reincarnated some 18
months ago.
In the new system a $50 camera photographs the produce from above so
as not to interfere with the countertop bar code scanner. The
algorithms were changed to do segmentation from the top, and the
system was reimplemented to run on a Windows platform.
As it is now configured, the system takes a picture of the vegetable inside its bag, which itself caused a problem because as Connell explains it, "The clear plastic bag is far from clear-it's kind of milky and it occludes the fruit underneath. It's difficult to see through, so we had to change the algorithms a fair amount to get rid of those problems." VeggieVision uses a combination of background subtraction and chromakeying to separate the vegetable from the superfluous information, such as the bag. It then looks at the color, texture, shape, and size of the vegetable and makes histograms of these features. The system concatenates all of this information into a feature vector that is about 300 bytes long. It then uses a "nearest neighbor matcher" to find the produce item in its memory that has a comparable vector and then registers the item. If it is not sure, the system will give the clerk a few similar items to choose from. The VeggieVision system can also be trained in that if it makes a repeated mistake and a knowledgeable clerk corrects it, it will learn what the produce item is. The system is now ready for prime time, and its developers have signed field test agreements with two scanner manufacturers and one company that makes self-checkout systems. Note: This project is also known as Veggie Vision (2 words). |
VeggieVision: A Produce Recognition System
R. M. Bolle, J. Connell, N. Haas, R. Mohan, G. Taubin
Proc. of the Third IEEE Workshop on Applications of Computer Vision (WACV-96),
Sarasota FL, pp. 224-251, December 1996.
More
| Contact: Jon Connell | Last updated: 6/12/02 | ||
 |
 |
 |
 |
