They’re the last item in your basket, apples. But when you’re using supermarket self-checkouts it can take forever trying to pick the right variety from the screen’s list. It’d be much easier if the machine could just identify them itself.
Toshiba’s system, developed by Susumu Kubota and his team at Toshiba’s research centre in Kawasaki, Japan, uses a webcam, image recognition and machine-learning software to identify loose goods, such as fruit. The company claims the system can tell apart products that look virtually identical, by picking up slight differences in colour and shape, or even faint markings on the surface.
When shoppers want to buy, say, apples at existing self-service checkouts they must choose the right product from a long list of pictures on a screen. Toshiba’s technology, part of which was presented last year at the 11th European Conference on Computer Vision in Chersonissos, Greece, compares the image captured by the webcam against a database of images and detailed information on the item’s appearance. The software uses an algorithm to produce a list of pictures of similar items, with its choice for the closest match at the top. If this choice is the correct one, the checkout user presses a button to confirm the purchase.
If the machine’s first choice isn’t the right one, perhaps because the angle of the previous photo was very different, the checkout user has the option to “re-educate” the system by choosing the correct product name to go with the picture of their item. The new photo then will appear at the top of the list.
“This system gets smarter as you use it more,” says Kubota. He says recent tests showed it was able to recognise produce even when it was placed in a clear plastic bag. Still, it is not perfect yet. Naoki Mukawa at Tokyo Denki University warns that users could take advantage of the re-educating mechanism to allow the wrong identification to go through because the mistaken product might be cheaper.
“Since Toshiba is thinking of implementing these to self-checkout stands, it would be interesting to know what they would do when shoppers are dishonest,” says Mukawa.
Kubota acknowledges there may still be holes in the system. “We are seriously looking at the problem of how to deal with misuse of the system by shoppers,” he says. “We might add a filtering step.”
Keiji Yanai of the University of Electro-Communications in Tokyo points to another difficulty. He says this type of object recognition system is more difficult to perfect than facial recognition technology, as it is harder to distinguish between generic objects. Similar ideas designed to identify products without barcodes have never made it to market in the past, he adds.
Toshiba says it hopes to commercialise the system within three years.
Image Credit: Pin Add