Evaluation of a Computer Vision Algorithm for Wood Defect Detection and Quality-Class Assignment of Birch (Betula spp.) Veneer Logs Under Industrial Conditions
Abstract
The quality grading of birch (Betula spp.) veneer logs, an economically important raw material in the Baltic region, is still largely manual, causing pricing inconsistencies, and automated grading under industrial conditions remains rarely studied. We evaluated deep-learning computer vision models for recognising wood defects and assigning a quality class on a sorting line. The algorithm and a scaler evaluated 153 logs using the same images. The recognition differed considerably between the defects: the central core-zone defects were recognised well (recall: core discolouration 97%, pith and drying cracks 91%, core rot 88%), while the dead and rotten knots and the less conspicuous defects were recognised poorly (knots 44%, ingrown bark 52%). The quality class assigned by the algorithm coincided with that of the scaler in 57% of the cases, and the algorithm tended to assign a lower class, mainly because of the over-indication of the core rot. For reference, manual grading itself was not fully repeatable: line scalers agreed with a detailed control scaler in 79% of the cases (69 to 88% between scalers). Automated grading is already accurate for core discolouration and pith and drying cracks, while knot detection remains the main limitation for industrial deployment.