|
Fabric defect detection is one of the key steps in the production process of textiles. In order to overcome the disadvantages of present visual inspection, which are low detection rate, low detection efficiency and heavy labor intensity, it’s particularly necessary to research the fabric defect detection technology based on computer vision, and it has important engineering significance.In this paper, based on the analysis and the comparison of existing defect detection theories and methods, the methods of fabric image pre-processing, eigenvalue extraction and the defect classification based on the Back-Propagation neural network for computer vision are analyzed and researched deeply.First of all, the noise characteristics and the noise source of fabric images are analyzed, a method that combines the median filtering with the wavelet denoising algorithm is presented, which has gained good denoising effect. Directed towards the problem that the image detail is blurred in the denoising process, the sharpening process to tone up the details is carried out by use of the Laplacian operator as the sharpening operator, which makes the image after pre-processing more sharp and easy for eigenvalue extraction.Secondly, a method to divide the image after pre-processing is presented by use of the periodicity of autocorrelation function, and a possible defect window were determined preliminarily according to the discrepancy between the window’s gray mean and the whole image’s average gray mean, then the window is made as the further detection area by a proliferation of Jiu-Gongge, which speeds up the speed of defect detection. The wavelet analysis algorithm is used to extract six eigenvalues of that area, namely the energy, the variance, the entropy, the difference, the contrast and the inverse difference features, as the basis of defect identification, which improves the defect detection accuracy obviously.Thirdly, a method to identify and classify defects based on 3-layer BP neural network is presented. The structure characteristics and the algorithm selection of the BP neural network are discussed thoroughly. By optimizing the structure of BP neural network, an optimization result of the number of the input layer neurons, the hidden layer neurons and the output layer neurons are presented.Finally, on the basis of the theoretical study, the tabby cloth is chosen as experimental object, and ten fabric samples containing lycra, buckle-off, warp-lacking, hole, unclean color, white pole, weft-lacking, miscellaneous fiber, yarn and non-defect fabric were inspected and analyzed respectively. The experimental results verify the feasibility and the validity of the proposed theoretical method.
|