Anomaly detection on arbitrarily distorted 2D patterns by computation of a virtual golden sample

This paper is concerned with the problem of anomaly detection on surfaces that show periodic decoration patterns as shown at car interior components for example. The decoration of the surface is allowed to show a smooth deformation that can vary from one sample to another. As a consequence, every sample is unique in its appearance. The task is to robustly distinguish between the regular decoration and possibly occurring surface defects. The proposed algorithm is based on the computation of a unique virtual golden sample that exactly matches the inspected part except from any defects. Through the virtual golden sample an anomaly image is computed that can be used for final defect detection and evaluation. The novel algorithm has been successfully tested for different types of periodic decorations patterns.