Finding defects in texture using regularity and local orientation

Abstract

We address basic aspects of detection of structural defects in regular and flow-like patterns (textures). Humans are able to find such defects without prior knowledge of the defect-free pattern. This capability to perceive local disorder has not attracted proper attention in machine vision, despite its obvious relation to various application areas, e.g., industrial texture inspection. Instead, numerous ad hoc techniques have been developed to locate particular sorts of defects for particular tasks. Although useful, these techniques do not help us understand the nature of structural defects, which is the primary goal of our study. In no attempt to compete with the existing dedicated algorithms, we approach texture defects based on two fundamental structural properties, regularity and local orientation (anisotropy). The two properties belong to a hierarchy of structural descriptions, with the former being a higher level one than the latter. Both properties have great perceptual value. In this study, they are assumed to underlie recognition of structural defects. Defects are viewed as inhomogeneities in regularity and orientation fields. Two distinct but conceptually-related approaches are presented. The first one defines structural defects as regions of abruptly falling regularity, the second one as perturbations in the dominant orientation. Both methods are general in the sense that each of them is applicable to a variety of patterns and defects. However, they are better suited to different kinds of patterns. Two tests are presented to assess and compare the two methods. In the first test, diverse textures are processed individually and defects are searched in each pattern. In the second test, classified defects in groups of textiles are considered. Conclusions concerning the scopes of the two approaches are drawn.