Abstract:

In order to improve the versatklity of detection algorithm on varying fabric defect types, an algorithm for woven fabric defect detection using dictionary learning (DL) is proposed. Firstly, normal fabric image is divided into small image patches and unfloded into a column vector, then all the column vectors are combined into a matrix. Secondly, the matrix composed of column vectors is solved by DL, and then the optimal dictionary(basis vectors) are extracted. Finally the dictionary is applied to reconstruct testing samples, where the reconstruction error between original and reconstruction image is used for defect detection. Effect of the patch size and number of dictionary is also investigated. Experiments on 4864 samples show that the proposed method can achieve 90% of detection rate with false detection rate below 10%.

Key words: non-negative dectionary learning, patch size, dictionary size, fabric defect, detection