Use Case
This use case focuses on automated visual inspection of products within industrial manufacturing processes. The goal is to detect defects such as moisture in dry materials, contamination, or insufficient mixing. A camera captures full-surface images of the products, and a trained algorithm evaluates whether the image matches the expected condition or indicates an anomaly.
A key challenge in industrial environments lies in their inherent complexity: not all potential defects can be clearly defined in advance. Therefore, the ability to detect previously unknown anomalies is especially important.
Since time is a critical resource in most production processes, this project focused on developing a solution that delivers precise results with minimal acquisition and evaluation effort—in real time and with high coverage.
The presented solution reliably detects anomalies as small as a pea on an area of 60×30 cm—based on just a single image per product. In the demonstrated example, sawdust was examined for unwanted materials and moisture spots. While many foreign objects are detectable in the visible spectral range, water usually remains invisible. In this case, moist areas can be clearly highlighted using infrared light—especially in the SWIR range—and made accessible for machine evaluation.
Complexity demands a deep learning approach
Classical image processing falls short for this task: production processes vary, training data on defects are often incomplete, and deviations are diverse.
Neural networks, especially deep learning models, offer the solution: they learn the features of the normal state from defect-free examples and later identify previously unknown anomalies independently.
During training, the network repeatedly processes the data, capturing patterns and deviations in a high-dimensional space. This ability to generalize makes deep learning the ideal method for companies seeking reliable quality assurance in dynamic production environments.
Why SWIR? Spectral advantages in the short-wave infrared range
The SWIR method exploits the absorption properties of water in the short-wave infrared range: moist areas appear with much higher contrast here than in conventional color images. This enables reliable detection of production defects that are barely visible to the naked eye.
Combining spectral data from multiple wavelengths with machine learning results in robust and efficient anomaly detection—a clear advantage that surpasses traditional methods in both precision and efficiency.
Camera setup: precise data acquisition
A multispectral camera system was used for data acquisition—a high-resolution color camera combined with SWIR sensors tuned to specific spectral bands in the infrared range. This configuration delivers consistent data cubes containing information across multiple wavelength ranges for each pixel. Precise calibration and synchronization of the sensors ensure reliable image quality for model training.
ML Network: Deep learning for non-deterministic scenarios
EfficientAD, a specialized autoencoder architecture, was trained for this use case. The advantage: only a dataset of defect-free training data is needed to learn the normal state of the production environment. Deviations from this pattern are later reliably detected as potential anomalies.
This method of anomaly detection reduces the need for comprehensive defect catalogs and is therefore ideal for companies with heterogeneous processes and variable surface structures.
Cuvis.AI: Platform for hyperspectral anomaly detection
Cubert offers Cuvis.AI, a platform specifically designed for developing, monitoring, and inferring anomaly detection based on hyperspectral data. It enables efficient training of machine learning models with spectral images and integration into industrial production lines—modular, scalable, and future-proof.
Dataset: High variability as a foundation
The dataset used consisted of multispectral images of sawdust with a wide variety of textures, colors, and spectral properties. This high variability ensures that the trained model can reliably detect deviations even in real production environments with complex surfaces.
Precise segmentation of the images allows for objective evaluation of the anomaly detection performance and continuous improvement.
Conclusion: Intelligent quality assurance for modern production
By combining hyperspectral imaging, spectral diversity across multiple wavelengths, advanced machine learning methods, and a robust platform, Cubert has developed a solution that elevates efficiency, precision, and automation in quality assurance to a new level.
For companies investing in Industry 4.0 processes, automated anomaly detection becomes a strategic tool: disruptions are detected early, equipment is better monitored, and machines operate more efficiently.
In short: With intelligent image processing and spectral anomaly detection, you ensure the quality of your production—reliably, scalably, and forward-looking.
About the Author
Dr. Matthias Locherer has been the Sales Director at Cubert GmbH since 2017. With a PhD in Earth Observation from Ludwig Maximilian University of Munich, he brings extensive expertise in remote sensing, spectral imaging, and data analysis. Matthias has contributed to numerous research projects and publications, particularly in the hyperspectral monitoring of biophysical and biochemical parameters using hyperspectral satellite missions. His deep knowledge of optical measurement techniques and physical modeling makes him a key driver in advancing innovative hyperspectral technologies at Cubert.
