Adaptive anomaly detection within near-regular milling textures

With an application to the quality control of steel, we present image processing algorithms for unsupervised detection of anomalies that are hidden within a global milling pattern. Thereby, we consider global Fourier-based approaches and the localized shearlet decomposition for damping the milling texture. These frequency-based approaches are compared to template-based autocorrelation in the spatial domain. As an alternative approach, the so-called matching pursuit is proposed with a joint Fourier and wavelet-based dictionary, whereby the two models for the global milling grooves and the localized anomalies can be simultaneously exploited. All these approaches are evaluated against a manually annotated real-world data set using quantitative and domain specific qualitative metrics.