Image inspection of knitted fabric defects using wavelet packets

Abstract

Image inspection by wavelet packets and a neural network classifier is presented for non-defect and six kinds of defects in knitted fabrics. The types of defect include a hole, set mark (coarse), dropped stitch, oil stain, streak, and tight end. In this study, wavelet packet decomposition of a sample image is carried out based on the best-basis wavelet packet tree with three resolution levels. The lowest-two entropy among all sub-band images and the standard deviation for the original image are selected as feature inputs of the neural network classifier. These textural features are shown in seven groups, which are separately distributed in the feature space. We gathered a total of 112 experimental samples, with 16 samples in each of the seven aforementioned categories. The results demonstrate that with the three features, 56 test samples are correctly inspected. However, the lack of one of the three features yields wrong classification of some samples. Therefore, the three features selected are definitely suitable for recognition of our knitted fabric defects and also are the smallest number of features required to give accurate inspection.