Surface defect detection is an important task of industrial inspection that has traditionally relied on trained human vision.
Automated and objective inspection methods based on image analysis have played a decisive role in the industrial
progress of the last decades. We propose a new unsupervised novelty detection method for defect segmentation in
textures. It uses a multiresolution Gabor filter scheme and shows the following properties: no need of any defect-free
references or a training stage; any adjustable parameters, and applicability to both random and periodic textures. We
apply the odd part of Gabor filters to the sample image, analyze the details obtained at different scales and orientations,
and extract a number of background texture features from the sample under inspection. In the analysis, we assume that
the wavelet coefficients of pixels can be suitably fitted by Gaussian mixtures, more specifically, by combining two
normal distributions. One of them would correspond to the background texture whereas the other would account for the
defective area. Since all the information is obtained from the sample image itself, the threshold selection is robust against
possible sample to sample fluctuations such as heterogeneities in the material, inplane positioning errors, scale variations
and lack of homogeneous illumination. The efficacy of the statistical analysis is demonstrated. The method is applied to
a variety of samples that exhibit either periodic or random texture. A comparison with other unsupervised method
designed for defect segmentation in periodic textures is done.