High-tech meets reality: When models struggle with real-world processes

At SAINT 2026, FOTEC presented current research results on AI-supported product recognition in industrial applications.

At SAINT 2026, Michael Kollegger and Kevin Janisch presented the results of a research project on the automated recognition of meat products using AI. The presented use case with Steirerfleisch (a type of Austrian meat) demonstrates that modern models can achieve high accuracy – however, data quality, stable recording conditions, and clearly defined processes are crucial for success.

This year, the Social Artificial Intelligence Night (SAINT) 2026 took place at the USTP – University of Applied Sciences St. Pölten (St. Pölten Campus).

At this event, Michael Kollegger and Kevin Janisch (FOTEC – a research company of the University of Applied Sciences Wiener Neustadt) presented insights into the practical application of AI in industrial environments. The Steirerfleisch use case was presented as a best-practice example.

The Use Case: Steirerfleisch

The Steirerfleisch project investigated how meat products can be automatically recognized during the ongoing production process. For this purpose, image data was collected directly at the plant over a period of one month.

Implementation of the AI ​​Solution

A CNN model (EfficientNetV2-B0) was used for product recognition. The image data was preprocessed and used for training and evaluation.

Results

The model showed:

• Good results with clearly distinguishable products

• Weaknesses with similar products and a small data set

• A strong dependence on the quality of the image captures

Key Insight

The model is not the only crucial factor; the following are paramount:

• Data quality

• Standardized capture processes

• Stable framework conditions