Fabric defect detection using adaptive dictionaries

In this paper, we present a new fabric defect detection algorithm based on learning an adaptive dictionary. Such a dictionary can efficiently represent columns of normal fabric images using a linear combination of its elements. Benefiting from the fact that defects on a fabric appear to be small in size, a dictionary can be learned directly from a testing image itself instead of a reference, allowing more flexibility to adapt to varying fabric textures. When modeling a test image using the learned dictionary, columns involving anomalies of the test image are likely to have larger reconstruction errors than normal ones. The anomalous regions (defects) can be easily enhanced in the residual image. Then, a simple threshold operation is able to segment the defective pixels from the residual image. To adapt more defects, especially some linear defects, we rotate the test image by a slight degree and re-analyze the rotated image. Compared to the Fourier method, experimental results on 47 real-world test images with defects reveal that our algorithm is able to adapt to varying fabric textures and exhibits more accurate defect detection.