Artificial intelligence (AI) allows for automation of inspection tasks in the confectionery industry that could not previously be solved by vision technology. The Berlin-based image processing specialist Bi-Ber has put an optical inspection system into operation at confectionery manufacturer Loacker, which relieves employees and ensures more reliable and, above all, 100 % quality assurance.

The Italian company A. Loacker Spa/AG is known for its crispy wafers and chocolate specialities and takes pride in selecting only the best natural ingredients from known sources. The family company is one of the market leaders in Italy and exports worldwide. Established in 1925, it now operates two modern plants in Italy and Austria.

Project Manager Engineering Markus Schönafinger emphasizes: “Sustainable raw materials management, controlled processes and proven quality have always been our guiding principles. As such, we strive always to be up to date with the latest technological advances.” For quality assurance, this means a continuous increase in the degree of automation. Where possible, the company re-places human-eye inspection with automated inspection systems to relieve employees. This commitment has now prompted the introduction of artificial intelligence (AI). The new, AI-based inspection system was developed and commissioned by the Berlin-based image processing company Bi-Ber GmbH & Co. Engineering KG.

The vision system checks chocolate wafers, which are still in the mould, for foreign objects and quality defects. Holes, breakage, smeared coatings and leaking fillings must be marked as rejects. Minor cosmetic flaws such as uneven coatings, air bubbles or scratches should ideally also be detected; however, there are greater tolerances for those. The AI evaluates individual chocolate products and assigns a quality value to each one. “Based on these scores and on the quality specifications, the operator can readjust the system during operation”, Markus Schönafinger notes. “The inspection system transmits a pass or fail signal for each individual position to the controller with a handshake so that rejected wafers are removed immediately after being turned from the mould.”

The wafers feature different shapes and sizes and chocolate coatings with varying cocoa contents. In addition, there can be a large variance in each product group – many optical deviations are permissible and are not considered defects.

Programming a conventional machine vision software would be too complex, since both the good pro-ducts and the defects can vary. AI lends itself to this task, because artificial neural networks learn to recognize good and bad products through sample images.

In this pilot project, Bi-Ber initially trained an AI for select products. Currently, the system is already in operation in quality assurance for one product and is still running in parallel to human-eye inspection for two other products. In the validation phase, employees check all products that the AI assesses as doubtful, and the system is adjusted accordingly. In the future, more than twenty product lines shall be tested by AI.

By now, various AI software modules are available, which “only” need to be trained for their specific task. Bi-Ber, a Cognex Partner System Integrator, has developed its first AI application using the Cognex VisionPro Deep Learning software suite. This software suite includes, among others, modules for segmentation, error detection and classification and may be trained using only good sample images. In contrast, Bi-Ber has opted for supervised training to accelerate the process and increase precision. The developers carried out tests with intentionally placed foreign objects, manually marked the image pixels and categorized error images.

“The inspection system has to segment and evaluate up to 54 frames per second,” reports Christopher Keiner, Software Developer at Bi-Ber. “The basic functionalities can be set up with deep learning in a manageable timeframe. It becomes challenging as soon as it comes to adapting the AI to the specific application. We spent the most time validating the results, optimizing the evaluation and adjusting the inspection speed to the production clock rate. However, these experiences have already a positive effect. The setup for the next products took place during production in four to five stages, in total about twelve hours per product. This is significantly faster and more efficient than programming a rule-based machine vision application.”

Bi-Ber tailored a system with a 17” panel PC housed in an air-conditioned stainless-steel enclosure. A compact, lightweight stainless-steel camera enclosure with a window pane is mounted directly above the transfer system. Mounted on the sides of the vision system, two beam lights with efficient, long-life LEDs make textured surfaces and foreign objects clearly visible. Two Gbit Ethernet cameras with 5 MP resolution each capture half of the chocolate moulds, which are moved onward every 2 s. The software merges the images and achieves almost distortion-free images of the full mould. The use of two cameras enables a greatly reduced working distance and thus an overall more compact design. The vision system features no moving parts and is extremely low-maintenance.

Christopher Keiner assesses the project: “This optical inspection application is virtually impossible without artificial intelligence. Admittedly, it took a lot of effort, but the learning curve for everyone involved was quite steep, and we would always recommend an implementation with AI again in the future for similarly demanding inspection tasks.” Markus Schönafinger sums up: “AI for quality control is a new technology for us. We took a deliberate risk with the investment. Bi-Ber has proven to be a strong partner who managed our high quality standards and large variety of products very well and built a stable, expandable solution.”

Bi-Ber will be exhibiting at ProSweets Cologne trade show (23 to 25 April 2023, passage hall 4/hall 5, booth C027).