Novelty detection on metallic surfaces by GMM learning in Gabor space

Defect detection on painted metallic surfaces is a challenging task in inspection due to the varying illuminative and reflective structure of the surface. This paper proposes a novelty detection scheme that models the defect-free surfaces by using Gaussian Mixture Models (GMMs) trained in Gabor space. It is shown that training using the texture representations obtained by Gabor filtering takes the advantage of multiscale analysis while reducing the computational complexity. Test results reported on defected metallic surfaces including pinhole, crater, hav, dust, scratch, and mound type of abnormalities demonstrate the superiority of developed integrated system with respect to the stand alone Gabor filtering as well as the spatial domain GMM classification.