Comparison of Various Feature Extractors and Classifiers in Wood Defect Detection

Detection of defects on wood during quality processes in the wood industry is extremely important both economically and in terms of production and use. In order to minimize the time and cost loss caused by products obtained with defective wood, manufacturers want to detect defects in wood early by applying quality control process. For this purpose, in this study, some experiments are carried out using texture analysis methods and machine learning classifiers to detect defective wood from wood images. The features of wood images in the dataset taken from literature are extracted separately with six texture feature extractors to detect defective wood. Features are classified using twelve different machine learning classifiers, primarily tree-based ensemble classifiers. Crossvalidation is used in all experiments to reduce classifier bias. The results obtained are presented comparatively in terms of each feature and classifier. The findings show that the most effective features in detecting defective wood are extracted by the Local Binary Pattern (LBP) method and the most effective classifier is the Random Forest Algorithm. An accuracy rate of 96.75 % is achieved with the LBP-RandomForestClassifier and, classification performance is also presented for each algorithm by creating hybrid feature vectors.