Traditional vs. AI-based Methods for Detection of Anomalies on Metal Surfaces

Anomalies are data that deviate significantly from expected patterns, often indicating defects or irregularities. Detecting such anomalies is especially important in manufacturing, where visual anomaly detection (VAD) based on image data plays an essential role in quality control.
By automating defect detection, VAD greatly reduces the time and costs associated with manual inspections. This has led to extensive research and the development of various approaches. While modern methods rely on deep learning (DL) techniques, earlier approaches are based on simpler statistical analyses.
This paper examines whether deep learning is necessary for detecting simple defects, such as scratches on metal surfaces. The results show that while AI-based methods achieve near-perfect detection accuracy, traditional methods using simple statistical features can still reach up to 89% accuracy. Additionally, these traditional approaches are far more efficient, requiring only a fraction of the inference time. This highlights their potential as a lightweight and effective solution, particularly in real-time or resource-constrained scenarios.

Keywords: Anomaly detection, Surface defects, Quality control, Image processing, Benchmarking.