Defect-driven quality classification of wooden lamellae using computer vision

The manual quality grading of wooden lamellae is a complex and time-consuming task, limiting productivity and consistency of industrial processes. An accurate automated sorting approach is crucial to stay competitive, optimize resource use, and adapt to future wood species. This study leverages computer vision to automate and improve quality sorting in a cost-efficient way. We collected a dataset of 5050 images from 2025 Norway spruce (<em>Picea abies</em>) lamellae captured using an industrial line scan camera. Each lamella was categorized into one of three quality classes: standard, industrial, or substandard. Based on this dataset, we developed and trained a Defect-Driven Quality Classification (DQC)—a hybrid model that uses defect segmentation and machine learning to perform the quality classification of the lamellae. An instance segmentation network (YOLOv8x) was trained to detect eight key defect types, including knots, cracks, and blue stain. To improve segmentation performance, we used cropped image patches, and a custom recombination algorithm to reconstruct full-length lamella predictions. From the resulting masks, we extracted quantitative features which were further used for classification. Additionally, we added texture features to capture the structural properties of wood. These features were used to train multiple classifiers—decision tree, random forest, gradient boosting, multilayer perceptron, and support vector machine—for quality prediction. The YOLOv8x model achieved a mean average precision ([email protected]) of 0.78. The gradient boosting model achieved the best classification performance, reaching 81% accuracy across the three quality classes. Our results demonstrate that hybrid vision-based systems like DQC can deliver reliable, interpretable solutions for industrial wood grading, with clear potential for improving product quality, production efficiency, and sustainability.