Vision-based inspection of industrial materials such as textile webs, paper, or wood requires the development of defect segmentation techniques based on texture analysis. In this work, a multi-channel filtering technique that imitates the early human vision process is applied to images captured on-line. This new approach uses Bernoulli's rule of combination for integrating images from different channels. Physical image size and yam impurities are used as key parameters for tuning the sensitivity of the proposed algorithm. Several real fabric samples along with the result of segmented defects are presented. The results achieved show that the developed algorithm is robust, scalable and computationally efficient for detection of local defects in textured materials.

No abstract

No abstract

No abstract

Automated visual inspection of patterned fabrics, rather than of plain and twill fabrics, has been increasingly focused on by our peers. The aim of this inspection is to detect, identify and locate any defects on a patterned fabric surface to maintain high quality control in manufacturing. This paper presents a novel Elo rating (ER) method to achieve defect detection in the spirit of sportsmanship, i.e., fair matches between partitions on an image. An image can be divided into partitions of standard size. With a start-up reference point, matches between various partitions are updated through an Elo point matrix. A partition with a light defect is regarded as a strong player who will always win, a defect-free partition is an average player with a tied result, and a partition with a dark defect is a weak player who will always lose.After finishing all matches, partitions with light defects accumulate high Elo points and partitions with dark defects accumulate low Elo points. Any partition with defects will be shown in the resultant thresholded image: a white resultant image corresponds to a light defect and a grey resultant image corresponds to a dark defect. The ER method was evaluated on databases of dot-patterned fabrics (110 defect-free and 120 defective images), star-patterned fabrics (30 defect-free and 26 defective images) and box-patterned fabrics (25 defect-free and 25 defective images). By comparing the resultant and ground-truth images, an overall detection success rate of 97.07% was achieved, which is comparable to the state-of-the-art methods.