A real-time system designed to detect and classify textile defects is presented. The system starts with an analysis of the optical Fourier transform of sample textiles. We also use a back-propagation neural network to help detect and classify defects. Experimental results show that the system is able to detect and classify nine out of the twelve kinds of defects in its data base.In 1989, Wood and Hodgson [ 4,5 ] used a computer to process carpet images and to classify the various kinds of defects on the carpets by autocorrelation. The method could not be used in the production line because it took a relatively long time to accomplish the job. In 1993, Ribolzi et al. [2] ] used an opto-electronic technique to detect defects in real time. First, they Fourier transformed the textile image and then examined its local peak values, which correspond to the zero-th order and first-order diffraction maxima of the twodimensional power spectrum. By comparing these zero-th and first-order maxima, obtained from both normal and defective textiles, they could detect defects. Although the method could detect textile defects in real time, it was able to detect only a few kinds of faults, probably because Ribolzi and his colleagues discarded the information in the area between the zero-th and first-order peaks.In this paper, we propose a procedure to detect and classify textile defects, which has the merit of both processes, that is, it is a real-time method and it is able to detect a number of defects. Similar to Ribolzi's method, our system starts with an analysis of the optical Fourier transform of the sample textile, but we consider two one-dimensional Fourier transforms derived from the 2D Fourier transform rather than the 2D Fourier transform itself. Furthermore, in addition to the diffraction maxima analysis, we also consider the region between the zero-th and first-order maxima. We think the features characteristic of each kind of defect are also present in this region because different spectra from different defects give rise to different fine structures in this region under close inspection. Probably because of this consideration, the method is able to detect more kinds of defects than Ribolzi's method.

It has been reported that textile defects can be found and classified by examining the optical diffraction pattern obtained by illuminating the defect-containing textile with a laser beam. Very little, however, can be found in the literature concerning the influence of the laser beam size on the performance of the method. We describe here a systematic method to determine an optimum beam size, and we demonstrate that the size of the illuminating beam must not be too small or too large, for the system to yield satisfactory results.

No abstract