Automatic feature selection for textile defects discrimination

In textile industry, defect detection is usually handled in an automatic way. In most cases, the occurrence of a defect in textiles is linked with a defect in the installation that is used to produce the textile. Defect recognition is done in most cases by human experts and it is an offline process. The human expert cannot work continuously more than an hour or two, without making mistakes. A rough estimation of the correctness of the recognition made by humans is 60-70%. This paper's aim is to present a methodology for feature selection that leads to high recognition rates and to simpler classification systems architectures comparing to the case when all the features computed in a given context are used.