Identification of Cotton Foreign Matter Using LINE-SCAN Based Hyperspetral Imaging System

Identification of Cotton Foreign Matter using Line-scan based Hyperspetral Imaging System

Yu Jiang

Changying Li

College of Engineering, the University of Georgia

Athens, GA

Abstract

Identification of cotton foreign matter is important to both the cotton and textile industry, because different foreign matter not only affects cotton’s monetary value but also causes various damages to textile products. Hyperspectral imaging technique has shown the capability of classifying foreign matter, but it provided a large amount of redundant information which limits the classification accuracy and processing speed. The goal of this study was to select the optimal wavelengths to be used for foreign matter classification. A total of 240 mean spectra were extracted from the hyperspectral images of cotton lint and 7 types foreign matter. Each type contained 30 replicates, and they were randomly separated into training and test set with the ratio of 15/15. The sequential forward selection method was applied on the dataset to select the wavelengths, which were used to reduce the dimensionality of original dataset as well as the optimal wavelengths for classification. Subsequently, the linear discriminant analysis with the wavelengths was performed to classify foreign matter categories with 2-fold cross validation. The experimental results showed that 22 out of 256 wavelengths were selected and they achieved classification accuracy of 93.33% or higher for all types. Therefore, the selected wavelengths could be used for online foreign matter classification systems.