Using clues to move paper coupons to mobile
The company, SnipSnap, which is now owned by Slyce, argues that it has digitized more than 250 million coupons over four years, for retailers that include Bed Bath & Beyond, Toys 'R' Us, and Aeropostale. It's learned the hard way to not get overly reliant on any one technique, even when working with coupon images from a retailer client, said SnipSnap President Ted Mann.
"There's a whole spectrum of different functions that we use to process an image," Mann said. They start with the easiest stuff, such as looking for a recognizable barcode or QR code. If the shopper takes a mobile image clear and tight enough to properly capture either of those two details, identifying and reproducing that image is easy. If the coupon happens to be from a retailer or consumer goods manufacturer client, the app can then do fun things like integrating that coupon into that company's mobile app and pouring the data into a CRM database. But before any of that can happen, that coupon has to be recognized.
For many of SnipSnap's customers, SnipSnap's coupon-recognition tactics are incorporated into that retailer's or manufacturer's app. But this puts a lot of the onus on shoppers. They must see the coupon, recognize it as coming from Retail XYZ, open that retailer's app (or perhaps even download it first) and then begin the coupon recognition process. Would it not make more sense for everyone to encourage shoppers to use a centralized app and to have the results then feed into the relevant app This is another good example of retailers deciding a strategy based on what is good for them — get everyone to use and stay within branded app — rather than what is easiest for their shoppers.
If it helps, doing what is easiest for shoppers is the most effective way to get them to ultimately do what you want them to do. If this is all about coupons that can only be redeemed by giving you money, isn't this a good time to back off and let the process that is most customer-centric take center stage
Oh, well. Such frustrations are part of life. Let's get back to coupon identification. SnipSnap customers give the vendor huge image libraries of all current coupons. In theory, this means that many coupons have to be directly identified as much as the coupon image needs to be checked against those databases and hopefully a match will be found. When that doesn’t work, Mann said, the system then searches the overall image for smaller images that it might be able to recognize, such as a company logo. At this stage, any hint is worth finding.
Also, this is an evolving process. Once the system identifies a coupon, it is far easier to quickly identify it the next time that coupon is seen. It’s like antivirus definitions software. Once someone somewhere gets infected, it should be identified and then every other instance of that virus is easily found and blocked. Although with a coupon, it's a case of being found and used. "When we have previously scanned a coupon, when we see that image, it's much faster," Mann said.
Speed is indeed crucial here, as the software has only several seconds to make its identification attempt.
The next effort is to examine what appears to be text and to analyze it using OCR (optical character recognition). Discovered text can be searched for, but an especially hopeful chunk of text would include an expiration date, which can be an especially helpful clue, Mann said. But OCR is not a perfect approach and it is especially vulnerable to the weaknesses of print. "It's not enough to just pick up the characters — which can be difficult if the coupon is crinkled or if the image is askew," he said. "You get some of it right and some of it wrong. The OCR piece is still the weakest link. It's the hardest to get the quality data just right."
Mann said that the most successful OCR efforts are the expiration dates and the coupon's face value (for example, $2 off).
The next step is an act of desperation, where the software forwards it to the lowliest of its subordinates: a human. "When a coupon comes in and we can't automatically identify all of those things, we kick it over to a human," Mann said, referring to a team of about 60 people working out of Nova Scotia. "A human goes in and tries to hand-tag the logo. We're using humans to train the system what to look for." For example, the person would literally mark the part of the image where he/she saw the identifying mark, such as a logo, to help the software learn how to identify more effectively. The worker "draws a boundary on that piece of the image and feeds into back into that particular node of the neural network," Mann said.
And even with a human, Mann said, the turnaround time can't be more than 10 to 15 seconds.
It looks like both the software and its human team are learning how to learn better. Now if only retailers could do the same.