Unsupervised defect detection in textiles based on Fourier analysis and wavelet shrinkage.

Abstract

An unsupervised approach for the inspection of defects in textiles by applying Fourier analysis and wavelet shrinkage is proposed. It does not rely on any reference images. For each sample under inspection, the periodic pattern in the background is first eliminated by zero-masking their dominant frequency components that show high gradient values in the spectrum. The Fourier-restored residual image is then denoised by wavelet shrinkage. The approximation coefficients and the processed wavelet coefficients are individually back-transformed to produce a pair of reconstructions from which either the low or the high-frequency information about the defects can be segmented using a simple thresholding process. The performance of the method has been extensively evaluated by a wide variety of samples with different defect types and texture backgrounds. The effectiveness of the proposed method is demonstrated by the experimental results in comparison with other methods.