This thesis examines the impact of temporal changes in the environment and on the objects themselves on methods for unsupervised surface
anomaly detection. It focuses on the so-called concept drift, which often
complicates fault detection in various contexts. The study delves into how
changes in image data—including illumination intensity, color temperature,
material reflectivity, the degree of non-uniformity of normal appearance (i.e.,
noise), and variously pronounced anomalies on objects—affect the performance of these methods. To investigate this, a synthetic image dataset was
generated using Blender, which allows complete control over all parameters
related to the aforementioned changes. The aim is to determine whether
pre-trained anomaly detection models can effectively handle minor yet gradual environmental shifts. To do so, the magnitude of each change is first
evaluated, and these values are then compared with the results produced by
selected anomaly detection models to explore any potential correlations.
|
