Deep learning-based wood surface defect detection using YOLO and RT-DETR with a low-cost real-time web application

Detecting defects on wood surfaces is a critical task in wood manufacturing, where imperfections such as knots, stains, and marrows directly affect product quality and economic value. This work systematically evaluates seven state-of-the-art deep learning architectures, YOLOv8n, YOLOv8x, YOLOv10, YOLOv11, YOLOv12, RT-DETR-L, and RT-DETR-X, for real-time detection across eight categories: Quartzite, Live-Knot, Marrow, Resin, Dead-Knot, Knot-with-Crack, Knot-Missing, and Crack. While recent YOLO (You Only Look Once) variants increasingly incorporate attention mechanisms to capture the global context, RT-DETR (Real-Time DEtection TRansformer) models rely on transformer-based encoder–decoder architectures for object detection. YOLOv8x achieves the strongest overall detection performance (recall: 0.74, mAP@50: 0.70, mAP@50–95: 0.41), while YOLOv10 provides the lowest inference latency (3.6 ms), closely followed by YOLOv8x (3.7 ms). The results indicate that convolutional feature extraction remains highly effective for capturing fine-grained texture variations, whereas attention-dominant architectures do not consistently improve detection under real-time constraints. To assess the feasibility of low-cost industrial hardware, the lightweight YOLOv8n model was further benchmarked on an NVIDIA Jetson Orin Nano edge device, achieving 13.92 FPS (Frames Per Second) end-to-end pipeline throughput. These findings reveal critical trade-offs between detection accuracy and computational efficiency, guiding the selection of suitable models based on industrial latency and accuracy requirements. To demonstrate practical applicability, a real-time web application is developed to enable automated defect detection in production environments. All data and code implementations developed in this work are made publicly available in the interests of transparency and reproducibility, and to support further research in industrial visual inspection.