Deep Learning-Based Computer Vision for Automated Wood Grading: Enabling Efficient and Adaptive Process Optimization

Wood is a natural, renewable material with excellent aesthetic and mechanical properties. Its increased use in construction can also support climate-change mitigation by acting as a long-term carbon store. However, industrial processing is challenging due to wood’s inherent heterogeneity. Climate change intensifies this problem by shifting forest composition toward a higher proportion of hardwoods, which increase heterogeneity even more. To address this, effective wood sorting is key, as it manages heterogeneity, boosts processing efficiency, and maximizes material use. At the same time, current automated sorting solutions remain costly and often difficult to integrate into existing processing lines due to space constraints. Consequently, many small sawmills still sort wood manually or not at all, underscoring the need for more accessible approaches. This thesis develops automated wood sorting systems using deep learning–based computer vision, providing a lightweight, cost-efficient and flexible alternative to existing market solutions. It covers multiple stages of the wood value chain, focusing on three sorting tasks: roundwood, lamella, and rod sorting. The first project starts at an early stage of the wood value chain, focusing on sorting roundwood by species and quality. By installing a camera system on an industrial sorting line, we showed that end-to-end Convolutional Neural Networks (CNNs) can effectively predict species and quality from visual data. However, deep neural networks are considered black boxes and are difficult for humans to interpret, which limits their industrial acceptance. To address this issue, we employed Explainable Artificial Intelligence (XAI) methods to evaluate model robustness. Furthermore, we developed the Segmentation Decision Tree (SegDT) approach. A novel concept of combining a You Only Look Once (YOLO)v8x instance segmentation model to detect defects with a decision tree classifier to predict the quality. This hybrid-interpretable method matches the predictive accuracy of end-to-end CNNs for quality assessment, while providing defect-level insights. Beyond species and surface quality, mechanical properties are crucial for construction. The second project moves further along the wood value chain, focusing on lamella. We captured lamella images with an industrial line-scan camera and labeled them with four-point bending test results. End-to-end models and an extended SegDT approach effectively predicted both quality and mechanical properties from these images. So far, we focused on standardized industrial processes with fixed dimensions, established techniques, and widely used species. The final project goes beyond current capabilities by sorting split wooden rods for novel hardwood boards. These rods have irregular shapes and variable dimensions, making the task significantly more challenging. Models predicted species reliably, while estimating mechanical properties remained challenging. Although deep neural networks extracted valuable information from the images, much of the variation remained unexplained. This thesis shows that deep learning–based computer vision offers significant potential for cost-efficient wood sorting. The approach is flexible and affordable, suitable even for small companies, and can be deployed on mobile devices, for example for in-forest sorting of logs. By lowering production costs and improving material yield, it can strengthen the role of wood in construction.