Speakers
Description
Reliable, high‑throughput evaluation of internal wood quality is a critical enabler for modern mechanized forestry and downstream processing. We report a compact set of results from an imaging‑based workflow for automated analysis of roundwood, focusing on (i) robust separation of bark from solid wood and (ii) consistent detection of slices containing internal defects such as knots, cracks, or degraded tissue.
Data and approach: We trained a three‑class semantic segmentation model on annotated log images with voxel dimensions of 1 × 1 × 10 mm. The training set was deliberately enriched with difficult cases, especially slices with discontinuous or partially peeled bark. To stabilize edge decisions with minimal computational overhead, we added a one‑step adaptive relabeling rule that converts specific high‑intensity peripheral pixels—frequently misinterpreted as background—into the bark class. For defect presence/absence, we introduced a composite multi‑layer representation by merging three consecutive slices into a single input image, thus injecting short‑range spatial context without resorting to full volumetric modeling.
Results — bark segmentation:. On a 50‑image validation subset, the Sørensen–Dice score for bark reached 0.704 (wood 0.981, periphery 0.996). Applying the single adaptive relabeling step increased bark Dice to 0.750, while maintaining wood at 0.981 and periphery at 0.999. The improvement held for both slices with defects and defect‑free slices, indicating that the rule targets a systematic boundary issue rather than overfitting to specific textures.
Results — defect detection: Single‑slice classification achieved 93.94% overall accuracy across logs. Introducing the three‑slice composite raised accuracy to 95.39% and increased the worst‑case per‑log performance above 92%, evidencing greater robustness to inter‑log variability. When the inter‑slice distance in the composite was increased to 2 cm, the accuracy reached 94.72%, suggesting diminishing returns once the composite exceeds the typical continuity scale of internal features. Class‑wise diagnostics further showed that pronounced dark regions associated with unhealthy knots, decay, and major cracks were detected with near‑certain confidence, whereas subtle, healthy‑knot transitions remained the primary challenge.
Implications for technology and operations: Three pragmatic interventions—targeted inclusion of difficult bark cases, a single adaptive post‑processing rule, and composite multi‑slice inputs—deliver measurable gains in segmentation reliability and defect detection while keeping the computational and operational footprint low. Such lightweight improvements can be integrated into mechanized inspection chains to support more dependable sorting, planning, and value recovery decisions in forest‑based value chains.
| Keywords | CT-imaging; deep-learning; bark-segmentation; defect-detection |
|---|
