Can AI grade birch veneer logs? An industrial test finds strong core-defect detection but weak knot recall

Introduction

Veneer log grading sits at the start of the plywood value chain, and in the Baltic region it is still largely a human task. A scaler standing at a sorting line looks at each log, judges knots, cracks, bark inclusions and core condition, and assigns a quality class within a few seconds. The classes drive the price paid for the log and the veneer yield that the mill can expect, so inconsistency between scalers or between shifts translates directly into money. A recent open-access paper in the journal Forests, written by three authors from the Latvian company VMF LATVIA SIA, asks a practical question: how well does a deep-learning computer vision system perform when it is tested on a real, operating sorting line rather than on a curated laboratory data set?

The paper is useful for sawmill and veneer-mill managers precisely because it does not stop at headline accuracy. It reports defect-by-defect recall and precision, a four-class agreement analysis against a human reference, and, importantly, a comparison with how well human line scalers agree with a control scaler. That last comparison gives a realistic yardstick: if people only agree with a control grader about four times in five, a machine cannot be judged against a perfect standard.

This article summarises the study, explains its numbers, and separates what the paper shows from what the portal considers a reasonable interpretation for mill practice.

Research methodology

Setting and data. The work was carried out on an operating roundwood sorting line at a commercial veneer mill in Latvia (the mill is not named). The logs were freshly cut, green birch (Betula spp.) with bark. Over time the system collected visual data on 1,673,568 logs. For the formal evaluation, the authors used a defect-containing subset of 153 logs. They note that this subset contains more reject material (18 percent) than normal production, where roughly 5 percent of shortlogs are rejects, so the test set is deliberately enriched with difficult cases. Control-measured logs in the study ranged from about 13 to 48 cm top diameter, 15 to 57 cm butt diameter and 2.5 to 3.3 m in length. A separate operational check compared line-scaler decisions with control-scaler assortments: 2,674 matched control assortments covering 5,341 shortlogs.

Imaging hardware. The side surface of each log is imaged by three colour line-scan cameras (left, right and above), triggered by an encoder. Together they cover about 75 percent of the surface; the underside is not imaged. The two log ends are captured by colour area-scan cameras (2464 x 2056 pixels) with 50 mm lenses. The cameras sit in an enclosed room with LED lighting. Inference ran on a workstation with an Intel Core i9-13900K processor, 32 GB RAM and an NVIDIA RTX 3060 graphics card.

Algorithm. The system uses convolutional neural networks implemented in PyTorch. Two U-Net segmentation models do the defect work: one segments defects on the end faces, the other finds knots on the side surface in image tiles. A separate EfficientNet model identifies species, but it was not part of this evaluation. Notably, the system does not predict the quality class directly. Instead, a rule-based module applies the requirements of the standard to the detected defects. Training used 1,033 end-face images and 417 knot images, split 80/15/5 percent into training, validation and test sets.

Quality classes. Grading followed the Latvian standard LVS 80-7:2024 for birch veneer logs, with Class I, Class II, Class III and non-conforming (reject) material. Seven defect types were evaluated: dead and rotten knots, pith and drying cracks, felling and bucking cracks, ingrown bark, dead-bark wounds, core discolouration and core (forest) rot. Class depends on defect size and on whether the defect reaches the peeling cylinder, the part of the log that is actually turned into veneer. As examples from the standard, as summarised in the paper: dead and rotten knots are not permitted in Class I, are allowed up to 40 mm in Class II and up to 80 mm in Class III; core rot is not permitted in Classes I and II and is allowed up to 50 mm in Class III; core discolouration is allowed up to 70 mm in Class I and is permitted in Classes II and III.

Reference and statistics. Training images were annotated by one experienced control scaler. The reference for the evaluation was a remote scaler (also an experienced control scaler, and one of the authors) who graded the same 153 images that the algorithm saw. The study assessed presence or absence of defects and the resulting class, not measured defect dimensions. For each defect the authors computed recall, precision and F1 against the remote scaler with 95 percent Wilson confidence intervals. For the quality class they reported the share of exact matches, a four-class confusion matrix, Cohen's kappa, and two tests of asymmetry (Bowker and Stuart-Maxwell).

Key findings

Core-zone defects are detected well. The table summarises recall and precision for each defect, as reported by the authors (n is the number of logs the remote scaler marked as having the defect).

The three rarest defects (core rot, dead-bark wounds, felling and bucking cracks) rest on very few logs, so their confidence intervals are wide; for core rot the recall interval runs from 53 to 98 percent.

Knots are the weak point. The algorithm found 44 percent of the dead and rotten knots that the remote scaler marked, meaning 56 percent were missed. Precision was better at 77 percent, and the false-positive rate on knot-free surfaces was about 7 percent. In other words, the main problem is missed knots, not false alarms. The authors attribute this mainly to birch bark, which makes knots hard to tell apart from the surrounding surface. They point out that the remote scaler looked at the same images and still found knots the model missed, so the shortfall lies with the algorithm rather than with the image quality alone.

Class agreement is only moderate. The algorithm assigned the same class as the remote scaler for 57 percent of logs (95 percent confidence interval 49 to 64 percent), with a Cohen's kappa of 0.40, which is conventionally read as fair agreement. When it disagreed it was more likely to be stricter: it assigned a lower class than the scaler for 26 percent of logs and a higher class for 17 percent. Both asymmetry tests were significant (Bowker chi-square 16.61, p = 0.005; Stuart-Maxwell chi-square 14.11, p = 0.003). The distributions illustrate the shift: the remote scaler's classes were I 23 percent, II 51 percent, III 8 percent and reject 18 percent, while the algorithm produced I 26, II 33, III 20 and reject 21 percent. The most common single movement was from Class II to Class III (20 logs). The algorithm rejected 32 logs, of which 11 were considered usable by the remote scaler (10 Class II, 1 Class III).

Core rot over-detection drives the downgrading. Of 20 core rot indications by the algorithm, 14 were not confirmed by the scaler and 7 were confirmed (these counts are as extracted from the paper and do not sum exactly; check the original). The authors suggest the model may confuse core rot with core discolouration and that rare classes are learned less stably. They also note that about a third of the lowered logs had no invented defect, so some downgrading may reflect a more severe measurement of real defects, which they did not test.

Humans are not perfectly repeatable either. Seven line scalers agreed with the control scaler at shortlog level in 79 percent of cases overall, with individual scalers ranging from 69 to 88 percent, each judged on roughly 200 to 1,400 shortlogs. The paper stresses that this figure is not directly comparable with the algorithm's 57 percent because the data sets and the unit of grading differ (shortlogs versus whole logs), but it shows that manual grading carries its own variation. The paper also reports control scaler effort of roughly 2 to 3 minutes per log (stated as not measured) against a few seconds for a line scaler decision; it does not report algorithm processing time.

Practical applications for sawmills

From the paper: The authors conclude that automated grading is feasible and already dependable for core discolouration and for pith and drying cracks, but that knot detection limits industrial use today. Their suggested route is to use the system as decision support for line scalers rather than as a full replacement. Their improvement list is concrete: add and improve training data for knots, explore combining colour images with 3D point-cloud data, image a larger share of the side surface, reduce core rot over-indication through distractor classes, better balanced rot data, texture descriptors and threshold tuning, measure defect size and position relative to the peeling cylinder more accurately, and evaluate grading per shortlog.

Portal interpretation (not claims of the paper):

• A hybrid workflow looks sensible. Core-zone flags from the camera system could be shown to the scaler as a prompt, while knot assessment stays with the human until recall improves. This matches the authors' own decision-support framing.
• Because the camera system seems to lean towards stricter classes, a mill adopting such a tool would want to track the financial effect of downgrading, for example how many logs flagged as lower class would have produced acceptable veneer.
• Mills should measure their own scaler-to-scaler agreement before judging any automated system. The 69 to 88 percent range among the line scalers in this study is a reminder that the baseline is not perfect.
• Any pilot should be run on production-representative log mixes. The paper's enriched test set means that class-level numbers cannot be read across to a mill's day-to-day flow.
• The same logic of imaging, rule-based grading from a standard, and human review may be relevant to sawlog sorting in sawmills, but this paper only studied birch veneer logs and offers no evidence for sawlogs.

Limitations

The authors list several limitations themselves. The reference came from a single remote scaler, so there is no estimate of how well a second expert would agree with that reference. That same person also annotated the training images, which may inflate agreement, although the authors describe the effect as small. Defect dimensions such as knot diameter or crack length were not evaluated, only presence or absence. The 153 logs were enriched with defects (18 percent rejects against about 5 percent in production), so precision and agreement do not transfer directly to normal production. Logs were graded as whole logs, not per shortlog, which is how the line actually sorts. Finally, the study covers one sorting line, one camera system and one species group.

Additional points arise from the data. The rare defect classes rest on only 7 to 11 logs, which makes their percentages uncertain, and the underside of the log (about a quarter of the surface) is not imaged. The full confusion matrix could not be read completely in the source used for this summary, so only the figures quoted above are given.

Conclusions

On a real sorting line, a deep-learning vision system recognised core discolouration and pith and drying cracks with recall of 97 and 91 percent, and core rot with recall of 88 percent but poor precision (33 percent). It struggled with knots (44 percent recall), ingrown bark (52 percent) and felling and bucking cracks (43 percent). The assigned quality class matched the reference in 57 percent of logs, with a systematic tendency to downgrade driven largely by core rot over-indication. Human line scalers matched a control scaler 79 percent of the time, so the benchmark is imperfect. The study supports using computer vision as decision support today and points to knot detection on bark-covered surfaces as the main technical hurdle.

Original research

• Authors: Janis Magaznieks, Mareks Millers, Agris Gabrans (VMF LATVIA SIA, Riga, Latvia)
• Title: Evaluation of a Computer Vision Algorithm for Wood Defect Detection and Quality-Class Assignment of Birch (Betula spp.) Veneer Logs Under Industrial Conditions
• Journal: Forests (MDPI), volume 17, issue 8, article 893
• Year: 2026
• DOI: 10.3390/f17080893
• Link: https://www.mdpi.com/1999-4907/17/8/893
• Licence: Creative Commons Attribution (CC BY; the page metadata states CC BY 3.0)

This article is an independent summary written by Sawmill Network Europe; the original paper should be consulted for full methods, tables and caveats.