Detection Of Fabric Defects On The Basis Of Visual Saliency Models And Support Vertical Machine

In the textile industry, defect inspection is a vital step for quality assurance in fabric production. Currently, the detection of defects in textile fabrics is usually performed by trained and experienced human inspectors. In recent decades, computer vision, image processing and pattern recognition technology is developing rapidly. As a result, the automatic detection of fabric defects, capable of producing higher-quality fabrics at a higher speed, is definitely desirable.In view of human can find a target with some features from the vision field quickly and exactly, an approach for detecting fabric defects based on visual saliency model and support vertical machine is present in this paper.Firstly, the visual saliency model is built to generate the saliency maps and the salient region is segmented. The composition of human visual system and the formation of visual attention mechanism are expounded. Then a model on the basis of Itti model is proposed for fabric defect detection. From this model, we get the final saliency maps. At last, Otsu method is adopted to threshold the saliency maps.After analyzing the grayscale distributions of the saliency maps with or without defects, the global salient associated values and the background salient associated values are extracted. The two features are then sent to SVM to judge whether there are any defects.To recognize the specific types of the defects, the features of the images should be extracted firstly. The features, such as geometry, texture and saliency features, are extracted in this thesis. Then the ratio of the average distances between the classes and among the classes is calculated as the performance indicator. With this ratio, the optimal combination of features can be got.At last, the fabric defects can be classified with a multi-class classifier. One-to-one strategy is chosen to expand SVM into a multi-class classifier. Then the optimal features, after being normalized, are sent into the multi-class classifier to realize the defect classification.The result of the experiments shows that the proposed algorithm based on visual saliency model and support vertical machine is effective and feasible.