We present a new approach to detecting defects in random textures which requires only very few defect free samples for unsupervised training. Each product image is divided into overlapping patches of various sizes. Then, density mixture models are applied to reduce groupings of patches to a number of textural exemplars, referred to here as texems, characterising the means and covariances of whole sets of image patches. The texems can be viewed as implicit representations of textural primitives. A multiscale approach is used to save computational costs. Finally, we perform novelty detection by applying the lower bound of normal samples likelihoods on the multiscale defect map of an image to localise defects.