Advanced image recognition research from the Advanced Digital Sciences Center (ADSC) has found an unusual testbed: a sushi buffet.

As pieces of sushi rush along a conveyor belt in Sushi Express in Singapore, an image recognition system quickly identifies each sushi roll and tallies the number of pieces per category. It helps employees know when they are running low on smoked salmon rolls, for example, and need to add more to the conveyor.

This system makes the sushi identification process seamless and simple, thanks to complex algorithms that were created by a team of researchers at ADSC, a University of Illinois research center in Singapore. The technology uses object detection and recognition techniques such as machine learning and classification algorithms.

“We were working with object detection and classification in the lab, and found that there was a need in industry that could also provide us with data to test our methods,” said Yuduo Zheng, a senior software engineer at ADSC. “We discovered that it can actually work well with sushi and provide useful statistics to our industry partner, Xjera Labs.”

Their system consists of two parts: detection and classification. To detect the sushi piece, the algorithm extracts what is called the local binary patterns (LBP) from the sushi—a type of visual descriptor in computer vision. The system then uses AdaBoost, short for adaptive boosting, which increases the algorithm strength and learning capabilities to better identify the correct piece of sushi.

They further improve their classification component with deep learning, which uses machine learning algorithms to extract data and make accurate predictions.

ADSC and Xjera Labs worked together to develop the system, and Xjera Labs has since licensed the technology and are involved in commercializing it for several chain sushi companies.

The ADSC team is continuing to improve the algorithms so the system will recognize new types of sushi, as well as seamlessly adapt to novel conditions and products of different brands and outlets.

“This technique could also be used for other conveyor-belt related companies and industries, such as logistics, by using it to check and count different types of objects on the conveyor belt,” said Zheng.

The ADSC team involved in this work included Zheng; Yongchao Wei, senior software engineer; Yue Xu, former engineer; and Lu Ding, software engineer.