It has been running since May 2024: the first sorting machine based on artificial intelligence in a Swiss production facility. Developed specifically for walnut sorting, it outperforms conventional systems that are five times larger in terms of efficiency. This breakthrough is the result of years of research and development, based on semantic segmentation and modern deep learning technologies.

Semantic segmentation is the current state of the art for the inline sorting of food. Each image is analysed at pixel level in order to classify material or categorise it into good/bad schemes. In addition to recognising defects, the technology also evaluates their size and significance. This is the only way to prevent tiny, insignificant irregularities from incorrectly rejecting an entire batch.

This method is standard in food processing. However, advances in deep learning are opening up even more far reaching possibilities.

The potential of deep learning solutions for food sorting is enormous. Anyone who has ever worked with such tools will be familiar with the impressive capabilities of these algorithms: They delineate objects precisely and classify them reliably. However, the implementation is often more complex than it first appears.

One key aspect is image capture. In inline processes, such as walnut sorting, objects are continuously transported on a conveyor belt and captured with line scan cameras. While this is unproblematic for classic semantic segmentation, neural networks require images in matrix form. These must be extracted from the image flow and calibrated in real time a challenge that requires enormous computing power.

The development of the neural network for the sorting machine presented the engineers with considerable challenges. Unlike conventional software solutions, a neural network does not follow any fixed rules. High quality initial data and a lot of time are crucial for success.

For the prototype, the team worked closely with a customer to manually sort large quantities of walnuts and use them as reference data. The training images had to clearly show the errors. Even the smallest ambiguities can confuse the algorithm and lead to faulty structures with a significant impact on the sorting quality. After years of development work, the time had come in May 2023: The first series machine was put into operation in Switzerland. It sorts walnuts faster and more precisely than conventional systems and marks a milestone in food processing.

This success illustrates the progress that is possible through the use of AI and deep learning. The combination of modern image processing, powerful algorithms and robust machine technology shows the potential that lies in the automation of processes.

The Strelen Control Systems team from Büttelborn (Germany) and Kreuzlingen (Switzerland) is proud of this milestone and is ready for new challenges. Because one thing is certain: The journey of AI-based sorting systems has only just begun.

This article is the prelude to issue 5-6/2025 of sweets processing. In the next issue, we will take an in-depth look at the use of AI in numerous areas of the value chain.