Towards Reliable AI YOLO Models for Automated Wood Defect Detection: Evaluation of Performance on Live Knots as a Subtle Class.

Global Climate action, supported by the Paris Agreement on Climate Change, has heightened the demand for environmentally friendly materials, such as wood. However, the wood industry is not fully braced to meet this demand because of production inefficiencies caused by reliance on outdated technologies and manual labor, for instance, in wood quality inspection. The availability of Computer Vision and Deep Learning models offers an opportunity to improve efficiency in the industry. While previous studies on real-time wood defect detection using YOLO models report higher performance. These reports mainly focus on overall accuracy scores, which mask the real performance of subtle defects studied, such as live knots, which are significantly important for realworld adoption of the tested models. By evaluating YOLO models' performance on both overall and individual defects, especially live knots, a hard-to-detect defect, this work seeks to address the gap in current research on YOLO models, fast-tracking their reliable adoption in the wood industry. To achieve this, a publicly available dataset was used to experiment with multiple YOLO variants, which were evaluated based on both overall detection accuracy and per-class performance. Initially, the YOLOv8n model served as the baseline for preliminary examinations. The dataset was preprocessed for computational efficiency, balanced across selected classes, and enriched through transfer learning. The evaluation emphasized metrics such as F1 score, precision, recall, and mean average precision (mAP) at the class level, with a particular focus on the poorly performing live knots. Additionally, for edge deployability, models’ Inference Speed and Model Size were also analyzed. From the evaluation, it was evident that while YOLO models maintain strong overall detection performance, they struggle with subtle defect classes. Some YOLO variants demonstrated improved robustness to class imbalance and subtle feature representation. The YOLOv9t model was the best at optimizing both performance and efficiency, highlighting its potential for deployment in defect detection. By shifting focus from aggregate performance metrics to per-class evaluation, this study provides insights into the practical challenges of automated wood defect detection. The findings highlight how differences in YOLO architectures impact not only detection performance on subtle defects, which are critical for real-time industrial applications, but also efficiency, as industries focus on Zero Defect Manufacturing (ZDM) practices. These results emphasize the importance of selecting appropriate model architectures for reliable deployment in the wood industry, thereby supporting its transition toward sustainable and efficient practices.