Wood Defect Inspection on Dead Knots and Pinholes Using YOLOv5x Algorithm

Abstract

Accurate detection of wood defects is crucial for ensuring the quality and reliability of wood pieces in various industries such as construction and furniture production. Some challenging defects such as dead knots and pinholes vary in size and shape, complex textures, and the presence of wood grains makes the inspection process more complex. Thus, this paper evaluates the performance of the YOLOv5x algorithm in detecting and localizing wood defects, especially dead knots, and pinholes. Using an augmented custom dataset and trained with transfer learning from a pre-trained model with COCO dataset, the algorithm achieves outstanding results. With a precision score of 96.5%, a recall score of 91.8%, and a mAP@0.5 score of 95.5%, it indicates highly accurate defect detection and localization. However, the model’s performance slightly dropped when considering mAP@0.5–0.95 which scored 66.1%, indicating challenges in detecting defects with higher IoU thresholds. Visual examples demonstrated the algorithm’s capabilities as well as instances of incorrect detections and failed detections. The findings of this study can contribute to the field of wood inspection systems and highlight areas for further improvement in defect detection algorithms.