OptiWood-YOLO: A Real-Time Defect Detection Method with a New Industrial Wood Dataset

Abstract

This study presents OptiWood-YOLO, a YOLOv8n-based pipeline for automated wood-surface defect detection. Defective timber materials were obtained in collaboration with wood-processing companies, and documented image acquisition included controlled laboratory imaging. The dataset comprises 10,507 images annotated for six wood-surface inspection categories—blistering, crack, fingerjoint, planing torn, scratch, and wrinkling—and one auxiliary Objects class. We investigate preprocessing and data augmentation and compare the resulting detector with YOLOv11, faster R-CNN, and RT-DETR under a fixed image-level train/validation/test partition. The reported model-comparison values for OptiWood-YOLOv8n are precision 0.54, recall 0.47, mAP@0.5 0.46, and mAP@0.5:0.95 0.23, compared with an mAP@0.5 of 0.45 for YOLOv11. These single-run point estimates indicate broadly comparable YOLO performance rather than statistically demonstrated superiority. The reported throughput for OptiWood-YOLOv8n was 51.30 FPS in the specified evaluation environment. Because the original image-level split does not guarantee independence of physical timber samples, and complete source and evaluation-provenance records are unavailable, the reported results should be interpreted as exploratory performance under the original partition, not as verified generalization to unseen timber samples or production lines.