<title>Invariant texture features for web defect detection and classification</title>

Abstract

Robust web inspection and defect detection requires the analysis of the local structure of the texture. This is of special importance if the inspection task is extended to different web types that vary significantly such as silk or wool cloth. We introduce an algorithm that combines local nonlinear invariant features with high discrimination capabilities and statistical classification. In addition, invariance with respect to Euclidean motion is crucial to industrial settings. Thus the features are based on the integration of nonlinear polynomials over the transformation group for which invariance is desired. This result in a feature vector for each pixel of the image that is invariant with respect to translation and rotation. Local texture variations that appear naturally in certain cloth types like wool therefore influence the feature space only partially depending on the design of the functions. Nonlinear functions have been shown to extend the feature space compared to linear functions. This improves the discrimination power of the feature with respect to other textile types. As a next step, the features are presented to a fully connected multi-layer perceptron network to classify the web defects. For network training, the error regions are manually marked on the original images and labeled according to the error classes. The images are divided into small patches for which the feature vectors are computed. To address non-error textural variations these regions are split up into several parts and trained separately. Experimental result son a database of 3200 textile images show a high separation capability of the invariant feature for classification of the defects.