The company, which today operates as GmbH & Co. Engineering KG, was founded in Berlin in 1997 by Ronald Krzywinski and others. It has focussed on a sub-area of quality assurance within confectionery production: The use of image recognition for defect identification. With its six employees, it has been operating from Wildau in Brandenburg since March 2023.

By Dr Jörg Häseler

The company sees itself as a system house for industrial image processing and has found its focus in the confectionery industry. It has its roots in the device technology graduates of Chemnitz University of Technology, although Krzywinski began his studies at a time when Chemnitz was still called Karl-Marx-Stadt. The company's headquarters was the inspiration for its name, as Bi-Ber meant nothing other than "Image Recognition Berlin". "The brand was set and the name has since established itself, because in addition to manufacturers, machine builders are also among our customers for the sophisticated camera systems," reports Krzywinski, ’in this way, more than 100 systems are in use worldwide for mould empty inspection alone."

The move into the confectionery industry was more of a coincidence, as the company had previously been involved in the testing of lead-based components, where the aim was to check whether the manufactured electronic parts were planar. The contact with confectionery then came about through an initial customer contact at a chocolate and caramel producer in Berlin, which still uses numerous Bi-Ber devices today.

"External capacity is used for the software development of customised image processing systems, design and electrical assembly. This means that the number of employees can rise to ten, especially in the creative phase with development, design, software development, assembly, adjustment, commissioning and service," explains the Managing Director.

The issues to be addressed include mould emptying control, the 3D inspection of bar products, the inspection of wafer sheets and, as a new approach, quality control with deep learning. The new AI-based technology in particular offers enormous advantages. Such a system is able to recognise errors in a short space of time. As a rule, the training phase of a deep learning system takes no longer than two days, whereby the manufacturer's expertise on possible errors must be gathered in advance.

A wide variety of moulds are used for the production of confectionery. The liquid mass is poured into these moulds. Before the empty moulds are refilled, it must be ensured that the moulds are free of product residues, fragments or splashes. Bi-Ber systems use colour image processing to check the moulds for these contaminants; there is no need for mechanical inspection using mould-specific stamps.

Using GigE or USB3 cameras with colour sensors and megapixel resolution, the vision systems detect residual contamination up to a size of 1 mm² in the alveoli of the empty mould. The cameras are connected to a panel PC with a Windows operating system. The cameras detect product residues based on the colour or the difference in brightness compared to the clean mould. An inverse evaluation to check for complete filling of the alveoli is also possible as an alternative. A digital or Ethernet interface is available for communication, as well as Profinet for connection to the higher-level process control system.

This technology makes it possible to automate tasks that are too complex for conventional image processing. Bi-Ber has implemented such a solution for numerous customers. The products to be inspected come in different sizes and shapes and are coated with different types of chocolate. The products are inspected inside the mould.

The inspection system must send an IO or NOK signal to the production system for each cavity. Depending on the product, the moulds contain up to 108 alveoli, which the inspection system must segment and evaluate. The inspection ensures that there are no pieces of plastic or metal on the products. Incompletely applied coatings, swelling fillings and breakage must also be ruled out. Less critical visual flaws such as uneven coatings should also be recognised.

The aim of introducing an AI-based automatic quality inspection system is to replace visual inspection and thus increase detection reliability. The final inspection in chocolate production is particularly challenging for a vision system because the products generally do not have a uniform appearance. Defects also always vary slightly. Conventional image processing works according to fixed rules that can hardly be programmed for this application, or only with considerable effort. With deep learning, on the other hand, artificial neural networks are fed sample images and use them to form their own criteria for categorising images. There can be a large variance in which products are rated as good - many visual deviations are permitted.

The deep learning application was developed with the help of Cognex VisionProViDi. This software suite contains various specialised tools, including ViDi RED Analyze for segmentation and error detection. Good and bad images are used to train the AI, and in supervised mode, NOK areas can also be marked in the images to focus on the defects being searched for from the outset. ViDi does not require huge image data sets and therefore saves a lot of time during training. The AI evaluates each product individually and assigns quality scores. The user can utilise these values to set the tolerance limit during operation and thus decide for themselves how homogeneous the products need to be.