Colour-patterned fabric-defect detection using unsupervised and memorial defect-free features

Automatic colour-patterned fabric-defect detection is essential and challenging in controlling manufacturing quality. Because of the scarcity of defective colour-patterned fabric samples and the imbalance of defect types, an auto-encoder trained with defect-free samples is used. However, the auto-encoder sometimes has weak generalisation ability, leading to mis- or over-detection of defects. An unsupervised and memorising defect-free method is proposed for colour-patterned fabric defects. The method designs a memory-guided quantisation variational auto-encoder-2 model and improves the residual post-processing operation. Specifically, it has three significant characteristics. First, it avoids time-consuming and laborious manually labelled samples and only needs defect-free samples in the training phase. Second, we expect the reconstruction image to be a defect-free image. Thus, memory modules are introduced to encourage the model to memorise the features of defect-free samples, which help to remove the image defect areas at the testing stage. Third, to further improve the detection accuracy, the closing operation is used to deal with the residual image that is calculated between the tested and the corresponding reconstructed image. Extensive experiments on various representative fabric samples demonstrated the effectiveness and superiority of the proposed method.