mRMR-based feature selection for classification of cotton foreign matter using hyperspectral imaging

A mRMR-based two stage framework was used for optimal wavelengths selection.A total of 12 wavelengths were selected for cotton foreign matter classification.The selected wavelengths were highly correlated to the properties of foreign matter.The selected wavelengths achieved 86% or higher hit rate for different classifiers. Different cotton foreign matter causes various levels of damage to textile products and decreases the monetary value of cotton. Hyperspectral imaging technique has shown the capability of discriminating the foreign matter, but its large amount of information which is mostly correlated and redundant limits the classification accuracy and processing speed. The goal of this study was to explore a new method of feature selection (minimum Redundancy Maximum Relevance algorithm) to select optimal wavelengths from the visible to near infrared spectra of the hyperspectral imaging data for cotton foreign matter classification. A spectral dataset containing 480 samples was collected from hyperspectral reflectance images of cotton lint and 15 types of foreign matter. Each sample was represented by a mean spectrum containing 256 wavelengths ranging from 400nm to 1000nm. The dataset was pre-processed by removing the noise, and the number of wavelengths was reduced from 256 to 223 by removing those with a signal to noise ratio lower than 10dB. The optimal wavelengths were selected from the pre-processed dataset by a two-stage approach. The first step was to rank the features using the minimum Redundancy Maximum Relevance algorithm and to provide only the top ranked features for the following feature selection. In the second step, the sequential backward elimination was applied to the top ranked wavelengths to select the optimal wavelengths for foreign matter classification. The generality of the selected wavelengths was evaluated by comparing the classification performance using the Linear Discriminant Analysis (LDA), Support Vector Machine (SVM), and Artificial Neural Networks (ANNs). A total of 12 wavelengths were selected as the optimal feature set for foreign matter classification. Eight wavelengths from the visible range were related to the natural or artificial pigments of foreign matter, and the other four from the near-infrared range were related to the proteins or nutrients in foreign matter. The selected wavelengths achieved average classification rates of 91.25%, 86.67%, and 86.67% for the LDA, SVM, and ANNs, respectively, indicating the generality of the selected features. This study explored a new method for hyperspectral imaging optimal wavelength selection and the selected wavelengths can be used with different classifiers for cotton foreign matter classification.