Data-driven textile flaw detection methods

Abstract

This research develops three efficient textile flaw detection methods to facilitate automated textile inspection for the textile-related industries. Their novelty lies in detecting flaws with knowledge directly extracted from textile images, unlike existing methods which detect flaws with empirically specified texture features.

The first two methods treat textile flaw detection as a texture classification problem, and consider that defect-free images of a textile fabric normally possess common latent images, called basis-images. The inner product of a basis-image and an image acquired from this fabric is a feature value of this fabric image. As the defect-free images are similar, their feature values gather in a cluster, whose boundary can be determined by using the feature values of known defect-free images. A fabric image is considered defect-free, if its feature values lie within this boundary. These methods extract the basis-images from known defect-free images in a training process, and require less consideration than existing methods on the degree of matching of a textile to the texture features specified for the textile. One method uses matrix singular value decomposition (SVD) to extract these basis-images containing the spatial relationship of pixels in rows or in columns. The alternative method uses tensor decomposition to find the relationship of pixels in both rows and columns within each training image and the common relationship of these training images. Tensor decomposition is found to be superior to matrix SVD in finding the basis-images needed to represent these defect-free images, because extracting and decomposing the tri-lateral relationship usually generates better basis-images.

The third method solves the textile flaw detection problem by means of texture segmentation, and is suitable for online detection because it does not require texture features specified by experience or found from known defect-free images. The method detects the presence of flaws by using the contrast between regions in the feature images of a textile image. These feature images are the output of a filter bank consisting of Gabor filters with scales and rotations. This method selects the feature image with maximal image contrast, and partitions this image into regions with morphological watershed transform to facilitate faster searching of defect-free regions and to remove isolated pixels with exceptional feature values. Regions with no flaws have similar statistics, e.g. similar means. Regions with significantly dissimilar statistics may contain flaws and are removed iteratively from the set which initially contains all regions. Removing regions uses the thresholds determined by using Neyman-Pearson criterion and updated along with the remaining regions in the set. This procedure continues until the set only contains defect-free regions. The occurrence of the removed regions indicates the presence of flaws whose extents are decided by pixel classification using the thresholds derived from the defect-free regions.

A prototype textile inspection system is built to demonstrate the automatic textile inspection process. The developed methods are proved reliable and effective by testing them with a variety of defective textile images. These methods also have several advantages, e.g. less empirical knowledge of textiles is needed for selecting texture features.