Automatic and non-destructive inner wood quality assessment of wood logs using 3D computed tomography and machine learning

Determining the quality of wood logs without cutting poses a significant challenge for boosting the productivity of wood suppliers and processors. Our study used 3D Computed Tomography scanning to inspect European Oak wood logs for inequalities. We scanned 26 oak logs with a resolution of 10 × 1 × 1 mm, producing an average of 390 to 430 images per log. These images were visually analysed and categorised into classes: Normal wood, Dead knot, Live knot, Rot, and Split. In the first phase, we investigated classification methods using machine learning for the two categories: non-defective and defective wood. In the second and third phases, we expanded our analysis to include all key defect classes: split, rot, dead and live knots. We utilised Support Vector Machines (SVM), Random Forests (RF), Decision Trees (DT), k-Nearest Neighbor (k-NN), AdaBoost, and Gradient Boosting (GB). In the initial classification (non-defective and defective wood), we achieved 90% and 93% testing accuracy with DT and RF. Then, we aimed to improve accuracy by replacing the feature extractors and using a bagging voting classifier. The third phase used two feature selection approaches, LASSO feature selection and Mutual Information, and we obtained 88%, 95%, and 90% testing accuracies.