Evaluation of feature extraction and selection techniques for the classification of wood defect images

The main objective is to evaluate different feature extraction and selection techniques as well as classification performances for the wood defect images. This study presents a classification system to classify the defect images from a database provided by a wood factory. This database consists of 1498 defect images and they are classified using Support Vector Machine (SVM), J48, random forest and K-NN classifiers. The features for each defect image are extracted using six types of feature extraction techniques. Feature selection methods are used to choose the features according to their significance. From the findings, it can be observed that Ranker method produced the best performance for most of the feature extraction techniques and classifiers. This directly indicates that all the extracted features have significant contribution. For SVM, it is tested with three different settings: linear, RBF and polynomial. The highest classification rate is obtained by using Gray Level Co-occurrence Matrix (GLCM) with SVM polynomial. For J48 and random forest classifier, features computed using Colour Coherence Vector (CCV) yielded the best measure, whilst for K-NN, it is Gabor features which performed best. Besides 89.85% of case crack are correctly classified, 38.63% for fungus, 16.48% for knot, 88.06% for worm holes and 51.61% for watermark case. For defect cases other than crack, it is observed that the number of misclassification cases is biased on crack case. The proposed methodology can be applied to create an automated visual inspection system for detection of semi-finished wood defect in the wood industry.