Fabric defect detection via learned dictionary-based
visual saliency
1. Introduction
Defect detection plays an important role in fabric quality control, as defects on the surface
will affect costs and grading of the final product. At present, inspection is still carried out by
humans in most of the production lines. However, manual detection method has a low detection
rate only up to 70 per cent, moreover, these manual operations are easily influenced by subjective
factors of workers. With the development of image processing and pattern recognition techniques,
automatic visual textile inspection systems based on these technologies can provide objective,
stable and reliable performance on defects examination. At present, many solutions have been
proposed, and they can be divided into three categories: statistical approaches, spectrum
approaches and model approach [1].
In statistical approach [2-4], spatial distribution of gray values is defined by various
representations, such as auto-correlation function, co-occurrence matrices, and fractal dimension.
And this method employs the statistics of background (defect-free regions) are stationary, can
effectively localize the defect region with distinct statistical behavior. However, different
statistical methods is suitable for the given fabric textures, and it cannot guarantee to adapt various
types of fabrics and defect. Spectrum approaches first transform texture images to the spectrum
domain using covers Fourier transform (FT) [5-6], wavelet transform (WT)[7-8], Gabor transform
(GT)[9-11] and filtering, and then apply some energy criterion for defect detection. Their
performance heavily depends on the chosen filter, and the choice of filter parameters is a quite
complicated task, since the detection performance heavily relies on how the filters can match or be
tuned to the property of a specific defect. Model approaches extract image texture features through
modelling and parameter estimation methods [12-13]. The defect detection problem can be treated
as a statistical hypothesis-testing problem on the statistics derived from this model. These methods
usually share a high computational complexity, but their detection results are not very satisfying.
Visual attention can enable biological and machine vision systems to quickly select the most
saliency object or regions from a scene[14,15]. Generally, fabric images always exhibit a high
periodic texture among sub-patterns. The defect will disrupt the local regularity (periodicity) of a
fabric, resulting in the defects are very outstanding among the fabric images with homogeneous
background. Therefore, visual saliency models will provide a promising method for fabric defect
detection.
Visual saliency models mainly consist of five categories[14]: information theoretic models,
graphical models, spectral analysis models, pattern classification models and feature-integration
models. Information theoretic models are based on the premise that localized saliency
computation serves to maximize information sampled from one’s environment. They deal with
selecting the most informative parts of a scene and discarding the rest[16-18]. Graphical model is
a probabilistic framework in which a graph denotes the conditional independence structure
between random variables. Attention models in this category treat eye movements as a time series.
Since there are hidden variables influencing the generation of eye movements, approaches like
Hidden Markov Models (HMM), Dynamic Bayesian Networks (DBN), and Conditional Random