Distributed Spacing Stochastic Feature Selection and its Application to Textile Classification

摘要

Many situations require the need to quickly and accurately locate dismounted individuals in a variety of environments. In conjunction with other dismount detection techniques, being able to detect and classify clothing (textiles) provides a more comprehensive and complete dismount characterization capability. Because textile classification depends on distinguishing between different material types, hyperspectral data, which consists of several hundred spectral channels sampled from a continuous electromagnetic spectrum, is used as a data source. However, a hyperspectral image generates vast amounts of information and can be computationally intractable to analyze. A primary means to reduce the computational complexity is to use feature selection to identify a reduced set of features that effectively represents a specific class. While many feature selection methods exist, applying them to continuous data results in closely clustered feature sets that offer little redundancy and fail in the presence of noise. This dissertation presents a novel feature selection method that limits feature redundancy and improves classification. This method uses a stochastic search algorithm in conjunction with a heuristic that combines measures of distance and dependence to select features. Comparison testing between the presented feature selection method and existing methods uses hyperspectral data and image wavelet decompositions. The presented method produces feature sets with an average correlation of 0.40-0.54. This is significantly lower than the 0.70-0.99 of the existing feature selection methods. In terms of classification accuracy, the feature sets produced outperform those of other methods, to a significance of 0.025, and show greater robustness under noise representative of a hyperspectral imaging system.