Anomaly Detection

The world around us is not perfect. However, most things work as we expect them and we build a mental model of what we consider normal. While detecting deviations from this norm in many cases is easily done by a human, machines struggle to do so. While modern machine learning systems achieved remarkable results in classifying objects, as long as there is sufficient training data available, many processes, especially in the manufacturing industries, are extremely well refined to produce the desired object, which from a machine learning perspective, unfortunately, leads to a large amount of normal data but limited defective data. This makes manufacturing processes very vulnerable against unforeseeable anomalies. 

We develop techniques that are trained purely with normal data but capable to detect, score, and localize anomalous data samples without ever having seen them. This results in more reliable and secure manufacturing processes.

The publications below show our recent advances in the field working with different data modalities, such as images and depth data.


Asymmetric Student-Teacher Networks for Industrial Anomaly Detection
Marco Rudolph, Tom Wehrbein, Bodo Rosenhahn, Bastian Wandt
WACV 2023

Fully Convolutional Cross-Scale-Flows for Image-based Defect Detection,
Marco Rudolph, Tom Wehrbein, Bodo Rosenhahn, Bastian Wandt
WACV 2022
[pdf] [code]

Same Same But DifferNet: Semi-Supervised Defect Detection with Normalizing Flows
Marco Rudolph, Bastian Wandt, Bodo Rosenhahn
WACV 2021
[pdf] [code]