Machine vision for birch boards: accuracy, economics and arithmetic checks

The paper by Ji, Zhang, Diao, Wang and Miao (Forests, 2023, Chinese Academy of Forestry) describes an integrated concept for birch (Betula) solid timber: species identification, CT-based log sawing optimisation, longitudinal sawing, quality detection and optimised cross-cutting. Only the detection and optimisation steps come with data. Region: China, tested at one factory in Shaowu City, Fujian, with timber stated to come from Latvia.

Set-up

• A line-scan camera (2048 pixels per line) with red stripe lighting at a fixed 60 degree angle, calibrated to about 0.1 mm per pixel.
• 4,000 boards with surface defects in 10 packages of 400; a typical board about 4000 x 58 x 43 mm, kiln-dried to 12-15% moisture.
• 13 defect types (for example bark, dead knot, crack, hole, mineral, wane).
• A convolutional neural network for detection. The paper names no specific network and gives no architecture, dataset size or training details.
• Cross-cutting solved as a dual-objective problem (yield and value); detection speed set to 150 or 214 m/min.

Reported results

Heartwood was recognised best (97.97%). Holes were mistaken for knots, dead and sound knots were confused, and light bark was mistaken for mineral. The paper gives no precision or recall and does not define "accuracy" formally. The weak classes are often the ones that set grade in appearance products.

Reported economics, and why to check them

For the test batch (33.56 m3) the paper reports an optimised value of USD 12,844.52 against USD 11,236.27 for manual marking, a difference of USD 1,608.25, presented as a 14.13% improvement. It also claims a decision speed of at least 60 m/min against 20 m/min for manual marking.

Worked example: checking the paper's own arithmetic

1. Difference: 12,844.52 - 11,236.27 = 1,608.25, which matches.
2. Relative gain: 1,608.25 / 11,236.27 = 0.1431, that is 14.31%, not the stated 14.13%.
3. Speed: 60 / 20 = 3 times faster decision speed, a claim that helps only if downstream equipment can use the extra throughput.
4. Per-board time: 0.311 + 0.368 = 0.679 s for scanning plus analysis, well under one second.

The portal found further inconsistencies: the listed volume terms sum to about 33.62 m3 rather than 33.56; the value-ratio percentages sum to about 122%; six value grades are mentioned but seven are calculated; production assumptions differ between sections (12,000 vs 1,000 m3 per month). The manual-marking comparison appears to use one low grade price for all manual output, which would bias the comparison. These checks were derived from the page text and not independently recalculated.

What to take away

• Detection near 90% for type and size and about 92% for location, with errors around 1% in dimensions, is a useful benchmark for expectations.
• The economic claims rest on one batch, one species, one factory and contain inconsistencies: indicative only.
• The CT-based log sawing step has no quantitative result in the paper.
• Authors acknowledge that detection errors remain.

Checklist for a vision-grading pilot

• Use your own species, drying quality and typical defects for training images.
• Report per-class results, especially holes, mineral and knot types.
• Use your own price list and a larger batch than 33.56 m3.
• Check that downstream equipment can use a speed gain.
• Recompute the vendor's or paper's percentages yourself.

Key points

• Birch board study (China): about 90% accuracy for defect type and size, 92.25% for location, missed-detection rate 0.084.
• Per-board scan plus analysis took 0.311 + 0.368 = 0.679 s.
• The reported 14.13% value gain recomputes to 14.31%, and other numbers are inconsistent: treat economics as indicative.
• Weak classes: holes, mineral, knot confusion.
• A speed advantage matters only if the whole line is balanced.