Breadcrumb navigation

Heterogeneous object recognition technology

Featured Technologies

March 5, 2018

Today, we are constantly hearing about the shortage of labor due to the declining working population, and the situation is particularly devastating on retail work floors at convenience stores and supermarkets. Of all retailing operations, payment operation (cash register operation) tends to have a heavier burden on clerks, and expectations are growing to make this an unmanned process.

The conventional automated payment system involves self-checkout machines using barcode scanning, but this still has issues in efficiency, such as the shoppers taking time to manually hold the barcode one by one over the scanner and not being able to handle non-barcoded items like produce. While an attempt to automate payment by attaching RFID tags to goods has begun, the cost of each RFID tag is high, which is preventing their full-scale introduction at general retail stores.

Meanwhile, NEC aimed for a technique that uses image recognition technology with a camera recognizing the retail product itself instead of applying barcodes and RFID. Through this technique, NEC realized unmanned settlement that only requires simple user processing and low operating costs on the introducing vendor. NEC's new heterogeneous object recognition technology significantly streamlined product scanning upon payment thanks to its accurate image recognition of any retailer product, regardless of how disorderly the products are placed.

Features of the technology

Products handled by retailers, such as convenience stores and supermarkets come in diverse forms: natural goods, such as produce, that have noticeable visible differences among the same type of individual items; industrial products that are very similar in design, as with packaged goods like beverages, snacks, instant noodles, and sundries; and products with the above characteristics combined, such as ready-made meals, delicatessen, and other daily-delivered goods where natural items come in packages. A uniformly accurate image recognition of such miscellaneous goods is near impossible, not to mention that recognition accuracy can be reduced in complex environments where numerous items are disorderly arranged, creating a major problem in practical application.

So we solved this issue by adopting the following two approaches:

  1. Integrating the two differently typified image recognition technologies based on deep learning and feature matching techniques respectively, the optimal combination of the two methods are used to maximize recognition accuracy for each product. This enables highly accurate recognition of various items ranging from natural objects to industrial products.
    Recognition of various items ranging from natural objects to industrial products
  2. By merely capturing the images of individual items, massive amounts of images of complex environment where numerous items co-exist in disorderly fashion are automatically synthesized for machine learning. This enables solid recognition of numerous individual items even when they are randomly placed, instead of having to physically capture many images of the actual complex environment.
    Solid recognition of numerous items in a complex environment where they are placed randomly

Future prospects

The unmanned payment system using image recognition technology is already being applied to advanced initiatives in regions where retailing is seeing rapid transformation into an IT service business, such as in the U.S. and China. NEC will achieve innovation in retailing through this heterogeneous object recognition technology.

Relevant Laboratories