Segmentation-based quantification of timber surface defects via improved YOLO-TDN and morphological measurement

This is the authors' abstract. We don't add key points for this paper.

This study proposed a non-destructive, computer-vision method for segmentation and size measurement of timber surface defects. An YOLO-TimberDefectNet (YOLO-TDN) was developed and coupled the segmentation results with morphological operations to obtain calibrated dimensions. A 506-image dataset covering three defect types (cracks, decay, holes) was constructed, and model hyperparameters were optimized via orthogonal and comparative experiments. To address small targets and complex wood textures, the spatial and channel reconstruction convolution (SCConv) module and the efficient multi-scale attention (EMA) module were introduced to enhance key-feature learning and context awareness. Ablation studies confirmed improvements over the baseline, with Box precision improving from 0.752–0.802 (+5.0%), Mask precision from 0.738–0.800 (+6.2%), and mAP@0.5 from 0.633–0.667 (+3.4%). Defect parameters relevant to structural performance – crack length/width and area/perimeter of decay and holes were accurately quantified using OpenCV algorithms with camera calibration. Robustness tests under varying illumination showed hole-area errors ≤ 3% and crack-width errors < 15% (mean 4.8%). The proposed pipeline provides camera-calibrated metrology for timber components, supporting in-service assessment, durability evaluation, and maintenance planning in wood engineering.