Image inspection of nonwoven defects using wavelet transforms and neural networks

Image inspection of nine kinds of nonwoven defects by the wavelet transform and neural network is presented. The defects include black yarn, hole, needle streak, oil stain, stripe, corrugation, white spot, folding mark, and wrinkle mark. The wavelet transform decomposes an original image into four subimages in different frequency bands. Four texture measures, the energy, contrast and correlation with gray-level co-occurrence matrices as well as the energy with wavelet coefficients, are selected as defect features and computed based on the low-frequency subimage at resolution level one. The feature values are distributed in groups by the categories of defects throughout the feature space, accounting for suitability of the four features for inspecting nonwoven defects. The subimage is a downsized approximation of the original image; thus, in this manner, feature extraction can not only consume less computation time but also maintain the classification accuracy. The neural network acts as a classifier, which is trained by forty-five samples. The experimental results demonstrate that among forty-five testing samples, the classification accuracy is 100 %.