Veneer log grading on an operating line

Most detection papers use curated images. The Latvian study by Magaznieks, Millers and Gabrans (Forests, 2026; VMF LATVIA SIA, the company that developed the evaluated system) tests a deep-learning system on an operating roundwood sorting line for birch veneer logs. It is valuable because it reports per-defect results and compares the machine with human scalers.

Set-up

• Three colour line-scan cameras image the side surface (about 75% of it; the underside is not imaged); area-scan cameras image the log ends.
• Two U-Net segmentation models find defects on the end faces and knots on the side. The system does not predict the quality class directly: a rule-based module applies the Latvian standard LVS 80-7:2024 (Class I, II, III and reject) to the detected defects.
• Evaluation set: 153 logs, deliberately enriched with defects (18% rejects against roughly 5% in normal production). The reference was one remote control scaler who also annotated the training images.

Defect-by-defect results

The core zone (the part that becomes veneer) is detected well. Knots are the weak point: 44% recall means 56% missed; the authors attribute this mainly to birch bark. The rare classes rest on 7 to 11 logs, so percentages there are uncertain.

Class agreement

The algorithm assigned the same class as the reference for 57% of logs (95% confidence interval 49 to 64%), Cohen's kappa 0.40 ("fair"). It lowered the class for 26% of logs and raised it for 17%. Of the 27 logs the reference rejected, the algorithm passed 6 as usable; of the 32 logs the algorithm rejected, 11 were usable by the reference.

Seven line scalers agreed with a control scaler at shortlog level in 79% of cases (individual range 69 to 88%). The paper stresses that this is not directly comparable with 57% (different data sets and unit of grading), but it shows that the human baseline is imperfect.

Worked example: rebuilding agreement from the confusion matrix

Rows are the reference class, columns the algorithm:

• Exact matches (diagonal): 26 + 35 + 5 + 21 = 87; 87 / 153 = 56.9%, the reported 57%.
• Lowered by the algorithm (above the diagonal): 6+3+0+20+10+1 = 40; 40 / 153 = 26.1%.
• Raised (below the diagonal): 13+0+7+0+3+3 = 26; 26 / 153 = 17.0%.
• Rejects passed as usable: 0 + 3 + 3 = 6.

Interpretation for mills

The authors recommend decision support for line scalers, not replacement. A sensible hybrid (portal interpretation): core-zone flags shown to the scaler as a prompt, knot assessment staying human until recall improves. Track the financial effect of downgrading, and measure your own scaler-to-scaler agreement first. The study covers one line, one camera system and one species group, and the data and models are not public; the paper offers no evidence for sawlogs.

Checklist for judging an industrial vision test

• Is the test set enriched compared with production?
• Who is the reference, and did they label the training data?
• Are rare classes backed by enough logs?
• Is there a human-versus-human baseline?
• Who funded or developed the system?

Key points

• Core discolouration and pith/drying cracks: recall 97% and 91%; dead and rotten knots: recall only 44%.
• Class agreement with the reference was 57% (kappa 0.40), with a tendency to downgrade; human line scalers agreed with a control scaler 79% of the time.
• Core rot precision was 33% (7 of 21 indications confirmed) on only 8 reference-positive logs.
• Test set was enriched (18% rejects vs about 5% in production), so numbers do not transfer directly.
• Authors work for the system developer; use the tool as decision support.