Automatic Fabric Defect Detection And Classification Using Machine Learning Technology

Fabric inspection is an effective assurance for the quality of textile products but requires great human job. It's a good choice to choose automated machine instead, which is faster and more accurate. In this paper, we investigate two aspects: fabric defect detection and classification. The main work is shown as following:1. Fabric defects have different shapes and this paper investigates the texture property of fabric defects using Gabor filters. Gabor filters can evaluate the target from different scales and orientations. This paper improves the traditional Gabor filters to make them better adjust the scale change of defects.2. We adopt the output of Gabor filters as feature and Gaussian mixture model (GMM) as the classifier. GMM can very well depict the distribution of unknown features. However, its order needs to be appointed manually. We use a method to evaluate the model order from training samples in addition to some available parameters. These acquired parameters can reveal the actual distribution of each fabric defect feature more accurately.3. We investigate Radial Basis Function (RBF) network as the classifier for defect classification. Traditionally, people usually use K-means method to cluster the training samples. One problem is, in such clusters, samples cannot be guaranteed to confirm to Gaussian distribution. We use GMM to improve it. After this, samples in each cluster are strictly Gaussian distribution, which conforms well to the parameters in radial basis function and can provide a better classification for fabric defects.4. We separately investigate the feature extraction method for fabric defect classification. We use a complex feature composed of Gabor and Local Binary Patterns (LBP) instead of using one single method. These two features can depict defect from global and local scales mutually. While this complex feature is of great dimension. We use Principal Component Analysis (PCA) and Generalized Discriminant Analysis (GDA) to process them. After this, the feature has fewer dimension, little correlation and between class distance, all of which are better for the training of Artificial Neural Network (ANN) classifier.