Research On Detection Method Of Defects In Fabrics Based On Machine Vision

Fabric defect detection is a key part of the on-line quality monitoring in textiles. In view of conventional defect detection methods, many problems such as the slow velocity, the low accuracy, and the high requirement of workers’experience exist, which can not adapt to modern mass production. In addition, most existing molding fabric monitoring equipment are abroad, the price of which are too high to be widely popularized. Thus it is in urgent need to develop an on-line fabric defect monitoring device with independent intellectual property rights.Multi-scale decomposition conforms to the requirement of human visual cortex on the effective representation of image, which approach to the optimal representation of two-dimensional singular signal. Compare with multi-scale wavelet transform and contourlet transform, nonsubsampled contourlet transform(NSCT) has the characteristics of multi-direction, translation invariance and sparse expression of the signal. This paper introduces the outline of the NSCT, and then puts forward an defect detection algorithm based on NSCT decomposition. First, the image is decomposed by the Gauss filter. After that the NSCT decomposition is implemented and a low frequency sub image and 8 high frequency band pass are obtained. For the low frequency sub image, non-linear gain and improved mathematical morphology methods are used to enhance its features. Then a global threshold is used to achieve defect pre-judgment and rough positioning. For the image judged to be defective, the two strongest direction response band-pass sub maps are picked to fuse. On this basis, this paper proposes a fabric defect detection method based on the NSCT and the function of standard deviation, which can eliminate the interference and capture the silhouette information of defect finely.To effectively classify fabric defects, both the local features and the global features are considered. This paper adopts a method combining the principal component analysis(PCA) and the BP neural network to recognize the defect types. In this process, a total of 14 characters are extracted from the LBP feature map through the gray symbiotic matrix (GLCM) method firstly. PCA is used to screening principal component data from the feature matrix. Finally, these principal component data are input to a well-trained BP neural network to determine the defects. In this paper,5 kinds of defects in textiles are used as experimental subjects, and the average classification accuracy of them is 95%.At the end, this paper designs a simple graphical user interface of fabric on-line defect detection system, which can visualize the algorithm flow and results intuitively.