Segmentation of defects in textile fabric

Abstract

For fast inspection of defects in textile fabric the complexity of calculations has to be reduced, in order to limit the system costs. Additionally, algorithms which are suitable for migration into hardware have to be chosen. Therefore, in this work a segmentation algorithm using first order statistics is applied. Preprocessing includes a logarithmic greyscale intransformation to obtain insensitivity to illumination changes. Afterwards texture features are extracted by a set of linear filters, which consider local neighbourhood relations. The filtered images are evaluated by histograms being calculated on a window grid. Finally, the histograms are classified by a Perceptron Net trained by Backpropagation. An interactive Teach-in program is provided to adapt the system to different kinds of textile fabric and appearances of defects

Automated fabric defect detection—A review

“…Yet, it produced distorted output for large-scale defects. No explicit results were presented in [95][96][97][98] and their reliability on a large database was not clear.…”

“…Its limitations [115] include its black-box character, difficulty in coping with abundance of features and concomitant variations in scale, position and orientation. Apart from those methods partially using NN in [14,19,39,48,71,74,82,95], NN-oriented fabric inspection methods are depicted in Table 5. Stojanovic et al [116] suggested a three-layer back-propagation artificial neural network for low cost fabric defect detection with off-the-shelf components.…”

“…Both are sensitive to noise [94] in the image. Neubauer [95] recommended a defect segmentation method based on multiple linear filters (including three separable convolution filters as 1st-order statistics). Depicting only one fabric sample in poor quality, the true positive rate and the true negative rate of detection were 98.3% and 90.6%, respectively.…”

Defect detection in textured materials using optimized filters

“…2) Optimization Using as the Object Function: The filter optimization with respect to the object function maximizes the squared difference of average feature values, normalized by the product of average feature values. The optimization with respect to entails the solution of using (23) and the chain rule of the differentation, the above equation yields (22) Substituting from (9) and (10) in 22, we get (23) where again 24The 23is the same eigenvalue problem as (21) in the previous optimization approach. Thus in this approach also, the eigenvector yielding the maximum object function is selected as the coefficients of the required optimal filter .…”

Textile Web Defect Inspection by Feature Analysis Method

“…A divided approach has been employed by Neubauer [7] to identify a defection on the fabrics. Meylani et al [5] have studied to apply two-dimensional lattice filter for the same objectives.…”