From Rule-Based to AI X-ray Inspection

Until now, X-ray inspection algorithms have been rule based – software extracts a series of measurements from image data and applies a set of rules to determine whether or not a product is conformant. In this situation, the engineer designing the algorithm will maintain an understanding of the underlying image processing and apply appropriate rules.

While this approach has worked well in the past, with the emergence of artificial intelligence (AI) and machine learning, it is beginning to look increasingly anachronistic. But to harness this much simpler approach, it is necessary to understand the processes of training systems.

Conventional wisdom would require that a large number of ‘classification’ samples are passed through the system, so that the equipment can ‘learn’ what is a good and a bad product. It is normally fairly easy to get, for example, 10,000 samples of good products to train the system. However, it is rather more of a challenge to get anyone in the food processing industry to make 10,000 reject samples!

The next step is therefore to create autonomous systems where we use the traditional rule-based algorithms to teach the machine learning system automatically. This has the advantage of deskilling the process and will result in systems that are much faster – both in terms of set-up and operation – as well as simpler to understand.

This article was published in the International Confectionery journal. Please follow this link to read more: