Selection of features for the classification of wood board defects

In this paper we compare three methods for selecting features that have recently been applied to the classification of defects on wood boards. A first method is based on statistical measures to determine how well features differentiate between classes. A second method consists of leaving out each of the features in turn and performing classification on the remaining features. A third method is based on genetic algorithms. The performances of the three methods are measured on a database containing color images of 900 pine wood defects classified into 9 categories. The best overall performance obtained was 93% of correct classifications on a test set, with only 20 out of 72 original features.