Research On Fabric Defect Recognition And Classification Algorithms Based On Wavelet Analysis And Support Vector Machine

Fabric quality control plays a very crucial role in the textile industry. Fabric defect is an important factor which affects the quality of fabric surface, Traditional fabric defect recognition and classification is mainly done by human inspection off-line, which is in heavy labor intensive, slow detection speed and low detection accuracy. With the rapid development of computer vision and pattern recognition technology, fabric defect automatic recognition and classification is an inevitable trend in production quality control of the textile. At present, research of fabric defect automatic recognition and classification has made some achievements, as image acquisition is easily affected by light change and noise and the defect is of many categories, fabric defect recognition and classification has been still a challenging research topic. Aiming at eight kinds of defect common in pure color textile cloth production, this thesis focused on fabric defect image preprocessing, the feature extraction algorithm, classification and recognition algorithm and the fabric quality automatic evaluation, its main work is as follows:First, this thesis built a database of fabric defect. Combined with the characteristics of the actual collection defect images, by setting an appropriate threshold, this thesis designed a method which could automatically select image preprocessing methods including histogram equalization and filtering and de-noising.Second, in view of the problems of poor smoothness of wavelet characteristic curve in the defect-free area by directly using the way of high-frequency fabric sub-graph window segmentation in which characteristic value extraction after the wavelet decomposition, this thesis designed a fabric defect feature extraction method based on wavelet coefficient difference. The horizontal and vertical high-frequency coefficient of the fabric defects in the third layer after wavelet decomposition respectively minus the coefficient of smooth reconstruction for the first time. The energy, entropy and variance were selected as the wavelet coefficient difference value of the fabric defect characteristics and then drew its characteristic curves. Experiment proved that the proposed algorithm was especially suitable for obvious texture fabric defect and flake defect feature extraction, and the proposed algorithm could fix a position on the defect effectively.Third, in view of the problem of high computational complexity and its solution lack of sparsity in the least squares support vector machine(LS-SVM) classifier, this thesis proposed a fast sparse approximating least squares support vector machine based on inverse fitting of fabric defect recognition and classification. FSALS-SVM was with Radial Basis Function kernel function, the area searching process of “grid” to find the(C,γ) parameters and "one-against-one" multiple classification algorithms classified eight categories of defect, and leave-one-out cross validation method ensured the robustness of the classification algorithm. Experimental results showed that correct classification rate of the mixed defect samples was 90.33% of the defect classification and recognition algorithm proposed in this thesis, and training time was shorter, its recognition and classification performance was better than that of least squares support vector machine classifier and BP neural network classifier.Finally, a computer automatic fabric defect grading method and fabrics automatic rating method based on image processing on the basis of the American four point defect scoring criteria of its transformation was designed. The recognition accuracy of this defect grading method of fabric defect can reach 2.8 mm. It’s more objective and efficient than the artificial defect evaluation.