Abstract Traditional wood identification relies on manual anatomical analysis, which is subjective, inefficient, and difficult to scale to meet modern high-efficiency demands. Despite rapid advancements in computer vision, existing wood identification methods predominantly operate as ‘black box’ models, often failing to provide diagnostic evidence consistent with standard wood anatomy criteria. Among hardwood anatomical features, vessels represent the most distinct and stable characteristics. This study develops a multi-stage deep learning pipeline for the automated identification and quantitative analysis of key International Association of Wood Anatomists (IAWA) vessel features, spanning a sequence of tasks from porosity classification and semantic segmentation to vessel grouping recognition and morphometric measurement. This study systematically compared candidate models at each stage to construct an optimal workflow. Experimental results demonstrate that MobileNetV3 excelled in porosity and groupings classification, achieving accuracy of 90.0 % and 96.2 %. U-Net achieved 0.939 mIoU in vessel segmentation. Furthermore, SPD-Conv YOLOv11 model attained an mAP 0.5–0.95 of 0.845 for vessel measurement, maintaining error rates for mean diameter and density at 6.5 % and 10.5 %. The system attained an 81.3 % accuracy in identifying IAWA codes based on 300 wood micrographs. This study presents an interpretable, automated framework for wood feature extraction.
Li et al. (Fri,) studied this question.