Computer vision-based prediction of quality and mechanical properties of wooden lamellae

Sorting wooden lamellae ensures high product quality, structural safety, and customer satisfaction. Key criteria include both visual features and mechanical properties. This study aims to develop a cost-efficient method for automatically sorting wooden lamellae. We present a dataset of 5050 spruce lamellae images with quality class labels from manual visual inspection. Each lamella undergoes a 4-point bending test to measure its modulus of elasticity (MOE) and modulus of rupture (MOR). Based on this dataset, two approaches classify lamellae and predict their mechanical properties. The first, the segmentation to properties (Seg2Prop) model, detects critical defects with a segmentation network and infers quality class and mechanical attributes from them. The second end-to-end approach predicts both quality class labels and mechanical properties (MOE, MOR) directly from raw images using either a residual network (ResNet)50-based convolutional neural network (CNN) or a vision transformer (ViT). Training of the models uses either original images or autoencoder-guided (AEG) representations. The autoencoder learns exclusively from defect-free lamellae, and its reconstruction errors highlight defects, guiding the model to the most relevant regions. Training and testing of the end-to-end models on beech lamellae evaluates their applicability to hardwoods. The Seg2Prop model achieves the highest classification F1-score of 0.81 and predicts MOE and MOR with coefficient of determination (<em>r²</em>) values of 0.58 and 0.59, respectively. Among the end-to-end models, the ViT with AEG inputs attains the best classification F1-score of 0.74. For hardwood boards, quality prediction reaches an F1-score of 0.54. Our approaches provide a cost-effective, efficient solution for predicting lamella quality and mechanical properties and demonstrate potential for adaptation to other wood species.