Feature evaluation and selection based on an entropy measure with data clustering
Hong Yan · Optical Engineering · 1995
We present a technique for evaluation and selection of features based on an entropy measure with 1-D K-means clustering of individual features. The technique compares favorably with a combinational selection technique based on the Euclidean distance separability measure in terms of computational requirement. Experiments on the handwritten numeral recognition problem using the multilayer perception classifier show that the technique can reliably evaluate features and successfully select those ones important for performing the classification using a system of reduced complexity with little degradation of the performance. The technique can also be used to discard the noise-corrupted features in order to increase the reliability of a classification system.