How a neural network saves a grocery store 52,000 rubles a month
There may be a queue for manual weighing
The convenience store has ~300 items to be weighed. When customers do this, a queue arises as they need to remember and enter the names or numbers of the items. When a cashier weighs, he spends up to two seconds looking for a hot item and up to 30 seconds looking for an unpopular one.
Federal trade network. Details are under the NDA.
The queue is an excuse to refuse a purchase, but additional cashiers are expensive
  • 40% of buyers are ready to refuse the goods if the line at the scale or checkout seems long to them.
  • If the queue is longer than six minutes, the buyer is ready to refuse a second visit to the store.
  • More cashiers are needed to increase the speed. More cashiers means more costs.
The principle of operation: the buyer or cashier puts the goods on the platform, and the system shows three products that, in her opinion, are on the scales. The most likely item is at the top, the least likely item is at the bottom. To reduce the number of errors and train the neural network, the final choice is made by the buyer or cashier.

Solution: smart scales that recognize goods by weight
Queues arise, among other things, due to the long search for the right product on the scales. We have prepared a prototype of a system that automatically recognizes weighted goods in a second.
Faster by 7 seconds
Previously, it took an average of 8 seconds to search for a product, now the neural network recognizes it in a second. The speed of identifying rare goods or products with complex names has increased by 15-30 times.

As a result
Decreased chance of fraud and error
When a person chooses a bulk product at a self-service checkout, he can choose a cheap analogue. When the cashier weighs, he is wrong in 2-8% of cases. The system reduces the chance of fraud and errors.

With such indicators, investments in the system pay off in 4-5 months

Reduced queues at cash registers, scales and self-service terminals
We calculated that a store with five checkouts saves 4-7 hours a day.

If we compare the savings with the release of an additional cashier or controller who reviews purchases at self-service checkouts, we get 52,500 per month

Minimum accuracy increased by 10 times
Average accuracy - 30%
With a dataset of 300+ photos, the accuracy has increased even more
Goods recognition accuracy
The more data, the more accurately the system recognizes the product. The system showed high accuracy with a volume of 200 photos of each product class.
100 photos
Minimum accuracy - 0.05
Average accuracy - 0.59
Maximum - 0.99
200 photos
Minimum accuracy - 0.57
Average accuracy - 0.89
Maximum - 1.0

Initially, we installed the system in two stores to determine the correct processor model, camera type, and camera positions.
Next, we integrated the development with the store's cash registers and trained the system to recognize fruits and vegetables.

Installed the system in seven stores to see if it could pay for itself and accurately identify products in high volumes.
As a result, for 200 commodity items, we received 160 photos each. The average accuracy of getting into a selection of three products was 94%.

Deployment prospects
The retail chain decided to install "Smart Scales" in 100 stores. The store's security service was also interested in the development, since the neural network helps to fight unscrupulous customers who take expensive weight goods, but choose a cheap analogue at the self-service checkout.

Technical nuances
The compute server is in the shop. This is necessary to protect data and speed up the system. Only updates come from the external server.

The system does not need the Internet, it is enough to connect to the network once a day to get the update.

We tested IP, WEB and a camera for computer vision. As a result, we chose an IP camera, because any installer can install it.

Equipment and resources
Other resources
We also needed a server, the work of our team and installation of the system.
Computing module
We chose from four options - Raspberry PI3, Nvidia Jetson TX2, Nvidia Jetson Xavier, Nvidia Jetson Nano. As a result, we settled on the Nvidia Jetson TX2, because its performance is enough to work at 3-5 checkouts.
  • The more photos, the fewer errors. But some products are rarely bought, so the dataset has to be collected manually.
  • The system may be wrong with similar products. For example, Granny Smith and Renet apples are similar in appearance. This was solved by displaying three probable positions on the screen, from which the cashier or the buyer chose the right one.
March 31, 2019
Prototype launch
May 31, 2019
The pilot is deployed in the federal chain of stores in Moscow.
Project Completion
September 1, 2019