Texture defect detection using wavelet transforms

Abstract

In recent years, achievements obtained in machine vision were on the way to facilitate computers replace several tasks heretofore performed by humans. Among those tasks, visual inspection of textile fabrics is one of importance. Inspection is performed by detecting and recording the position of defected parts in a roll of textile web. Textile fabric images carry textural properties. Therefore, texture analysis methods can be incorporated to device algorithms for the solution of the defect detection problem. Wavelet transforms have proven to constitute powerful means, suitable for several image processing applications. In this thesis, wavelet transform based feature extraction methods are investigated in detail. Pyramid structured wavelet transform (PSWT), wavelet packet (WP) expansion and multichannel features are compared with features derived from spatial domain co-occurrence matrices in terms of defect detection capacity. Finally, a novel feature extraction scheme called subband domain co-occurrence matrices is proposed and compared computationally and performance-wise with the rest.