Anomaly detection is a method in which an automated identification of unusual or deviating patterns in data, processes, or surfaces occurs. In industrial environments, an AI-based method is frequently used for this purpose, as it enables automated surface inspection where significant deviations from the expected normal state are detected.

Unlike other AI approaches where specific defect types must be defined in advance, anomaly detection using machine learning does not require predefined defect classes, which simplifies implementation.

Goal: Classification or regression based on annotated data

Common use cases in industrial quality control:

• Segmentation (e.g. defect localization on surfaces)

• Classification (e.g. product type A vs. type B)

• Object detection (e.g. verification of the presence or correct positioning of components)

• Optical Character Recognition (OCR) for reading labels, serial numbers, etc.

Goal: Detecting patterns or structures in unlabeled data

Common use cases in industrial quality control:

• Anomaly detection (e.g. identifying abnormal image or process patterns)


• Clustering of process data (e.g. root cause analysis)


• Feature reduction (e.g. dimension reduction before model training or visualization)

An anomaly detection AI model is trained exclusively on defect-free data points. The algorithm develops an internal representation of the appearance of the defect-free state. Specifically, this means that the model learns what is considered "normal". AI-based methods are capable of precisely capturing even complex normal states, thereby avoiding unnecessary false alarms due to non-critical deviations within the OK range. Deviations from this learned normal state are detected as statistical outliers during operation.

Anomaly detection with machine learning offers a quick entry into automated quality inspection because it does not require time-consuming defect annotation. In industrial applications, however, the algorithmic process often reaches its limits – especially when high precision or detailed defect information is required.

Anomaly detection and segmentation models each have specific strengths – and complement each other perfectly when it comes to ensuring reliable and efficient quality control in production.

While anomaly detection operates without complex defect annotation and provides early warnings of new or unknown deviations, segmentation, classification, or object detection models provide more precise information on the type and location of specific defects.

At Maddox AI, we strategically combine both approaches:

• Primary inspection tasks such as defect localization and classification are handled by dedicated segmentation or classification models.

• Anomaly detection runs continuously in the background and alerts users to new or previously unseen defect patterns.

This ensures that no relevant outliers or deviations go undetected, even if products, processes, or defect types change over time.