A wrapper-based feature selection approach using Bees Algorithm for a wood defect classification system

Identifying defects and classifying them according to some predefined classes is common in many manufacturing processes. The basis of such approach depends on a set of features extracted from all the classes and using them to train a classifier and then use it to determine the class to which the unseen data belongs to, with a reasonable accuracy. Hence the performance of the classifier depends on the features´ ability to discriminate between the good or normal and the defects. Therefore, one way of improving the classifier is to select the most appropriate features from a given feature set for the purpose of training and testing so that, at the end, better results can be achieved overall. In this paper, a novel wrapper-based feature selection approach using Bees Algorithm for the application of wood defect classification is presented. Bees Algorithm is a swarm-based optimisation technique mimicking the foraging behaviour of honey bees found in nature. In order to demonstrate the wrapper-based feature selection procedure a Minimum Distance Classifier (MDC) is used in this study. However, the method can be applied to any application using some other classifier. The study shows that, on average, a 10% improvement is achieved when a reduced sub-set of 12 features selected using the proposed wrapper-based method with Bees Algorithm is used in training and testing the MDC when compared to using the original full set of 17 features. The rejected features correspond to outliers.