Fabric defect segmentation using multichannel blob detectors

The problem of automated defect detection in textured materials is investigated. A new algorithm based on multichannel filtering is presented. The texture features are extracted by filtering the acquired image using a filter bank consisting of a number of real Gabor functions, with multiple narrow spatial frequency and orientation channels. For each image, we propose the use of image fusion to multiplex the information from sixteen different channels obtained in four orientations. Adaptive degrees of thresholding and the associated effect on sensitivity to material impurities are discussed. This algorithm realizes large computational savings over the previous approaches and enables high-quality real-time defect detection. The performance of this algorithm has been tested thoroughly on real fabric defects, and experimental results have confirmed the usefulness of the approach.