Abstract:

The conventional cenral local binarization mode (CLBP) used in fabric defect inspection has the problems of high histogram dimension and feature redundancy, and limitation exists in conventional CLBP when the amplitude of the small part of the image varies greatly or the amplitude is flat. To solve the problems, a modified discriminant complete local binary pattern with lattice segmentation for fabric defect inspection was proposed. The proposed algorithm was divided into two a training part and testing part. The training stage was to calculate the feature value for each lattice after lattice segmentation in defect-free images and acquire the mean value of all feature values. The threshold was calculated by calculating the relative divergence between the feature value of every lattice and the mean of the feature values. The testing stage was to calculate the relative divergence and compare the result with the threshold. The lattice whose result was larger than the threshold was marked as a defect area. The proposed algorithm was compared with local binary patterns, boolean line indicator method, regular band method algorithms. Testing on fabric image datasets including 2 kinds of textures and 3 kinds of defects shows that the method has better inspection effect on star pattern and box pattern fabrics, one part of the positive rate (TPR) value can reach 0.99, and most of the inspection results of TPR are above 0.90.

Key words: central local binarization mode, lattice segmentation, feature extraction, relativedivergence, fabric defect inspection