Working with HES-SO Valais-Wallis’s Institute of Informatics, the Swiss group Bühler has developed an AI algorithm that improves the efficiency, reproducibility and speed of quality control methods in the food production chain. With the help of Valais expertise in the field of continual learning, the Institute has devised a unique algorithm that includes human experience. The result? Improved uniformity in industrial quality control processes.
Rice, a staple food for more half of the world’s population, must be correctly sorted and cleaned before it is distributed on the market. The Bühler group, with a workforce of 13,000 in more than 140 countries, specialises in mechanical engineering and thermal processes. It has developed mechanical cleaners and optical sorters capable of sorting up to 20 tonnes of rice an hour, using a jet of air to expel defective rice grains.
The project, which was begun in early 2017, involved the automatic analysis of millions of grains of rice in a bid to improve the accuracy of quality control.
Adding human input to the continual learning loop
The algorithms developed in the first stage of the project produced promising results with an accuracy rate of 85%, a welcome outcome but not sufficiently reliable for industrial applications. The scientists at HES-SO Valais-Wallis and the team at Bühler therefore decided to integrate human experience through a process of continual learning.
To improve the overall quality of the process of sorting the acceptable grains of rice from the substandard ones, they developed a loop training process. This stage resulted in a web-based app that the industrial client can use to load new images, monitor how the classification process is performing and correct any errors, producing an overall accuracy rate of more than 99%.
