Anomaly Detection in a real industrial use-case using Deep Learning

Anomaly detection in industrial contexts is a critical component of quality control. Traditional methods for performing this task rely on manual inspection or classical computer vision techniques, which are time-consuming and require careful tuning of many parameters to work in one specific case. In recent years, deep learning-based methods for identifying anomalies in images have grown in popularity. One of the goals of this thesis is to study the state-of-the-art and experiment to see if these methods can be used in real-world scenarios. In particular, the analysis is focused on a challenging dataset composed of images of aluminum molds to search for cold junctions, a class of anomalies difficult to detect even for experts. The two main techniques analyzed are PatchCore and Reverse Distillation. Both showed mixed results with respect to the dataset, correctly identifying the clearest anomalies but failing to identify the most challenging ones.