Optimized feature selection for improved tropical wood species recognition system

Abstract

An automated wood recognition system can classify wood species in just a matter of seconds and can replace manual inspection by human. The system captures image of wood surface and features are extracted from these images before being input into the classifier. However, due to variations in tropical wood species, a single feature extraction method cannot accommodate multispecies classification. Therefore, we propose the use of two feature extraction method which are Basic Grey Level Aura Matrix (BGLAM) and statistical properties of pores distribution (SPPD). Features from both feature extractors are fused together. However, the use of the fusion method resulted in the increase in the number of features for classification. In order to ensure only features that are discriminating enough as well as reduce the number of features, we developed an optimize feature selection algorithm to the fused feature set by selecting only discriminant features to be used for classification. A wrapper Genetic Algorithm (GA) is introduced into the system to make feature selection and classification. Linear Discriminant Analysis (LDA) is wrapped inside GA to select only discriminating and non-noisy features to be subset database for final classification. k-Nearest Neighbour (k-NN) classifier is also used for comparison. The result of experiments shows that the combined features of BGLAM and SPPD give good classification results. Wrapper based GA can select good features and increases system's final classification accuracy.