Does AI improve yield in secondary wood processing?

Against fixed-length cutting, yes, in the one recent trial that measured it. A 2024 study ran 100 short, dry boards through a cutting system guided by an AI defect model, and yield rose from 69.2% with fixed-length cutting to 81.5%. The paper aims at small plants that still mark defects by hand. Against a plant's existing scanner and optimiser, we found no published comparison. And an older test shows how vision can lose wood. In 1998, an automated system yielded 56.3% against 65.6% in the actual rough mill, because it "tended to identify and cut out defects that were not truly present".

What the usual answer says

The common answer: computer vision detects defects and optimises cutting, raising recovery and reducing waste in cross-cutting, ripping and finger-jointing. High detection accuracies are often quoted. The step from detection to yield is where the evidence thins.

What the 2024 study measured

Fan, Zhuang, Liu, Yang, Zhou and Wang, in Sensors, trained defect models on 1,050 images and tested on 450. Their chosen model reached a mean average precision of 0.94 across four defect types: dead knots, live knots, cracks and pith. They then built cutting rules that place saw points around detected defects.

The test used "dry, finished sawn timber with a size of 1200 mm × 150 mm × 20 mm" of five species, bought on the open market. "Using a conventional fixed-length truncation approach, the yield amounted to 69.2%." The AI-guided scheme "yielded a remarkable 81.5%". The authors call this "a 12.3% increase in wood utilization rates compared to fixed-length processing". Strictly, it is 12.3 percentage points, or about 18% more usable wood, on short uniform boards in a controlled run.

The paper is clear about its audience. "Many small-scale wood processing facilities exhibit relatively old processing methods, often relying on manual marking and sorting." For such a plant, fixed-length cutting is a fair baseline.

The authors also show where the model went wrong. "A live knot was misidentified as a dead knot because of its excessively silver appearance", and "a shadow along the edge of the lumber was erroneously labeled as a dead knot". A dense patch of grain was classed as a live knot. The paper doesn't report what those errors cost in yield.

What the 1998 test showed

An older study answers that question for its time. Kline, Widoyoko, Wiedenbeck and Araman, writing in Forest Products Journal in 1998, used 134 red oak boards to compare gang-rip-first yield from a prototype colour-camera inspection system with the estimated optimum and the yield actually measured in the rough mill. "Automated yield was found to be 56.3 percent compared to 69.1 percent (optimum) and 65.6 percent (observed)." The reason: the algorithms "were very sensitive and tended to identify and cut out defects that were not truly present".

The technology has changed a great deal since. The lesson hasn't. A shadow read as a dead knot, as in the 2024 paper, is the same kind of error, and every false defect removes clear wood.

Detection scores are not yield

A 2025 preprint by Kang, Cen, Cen, Wang and Liu built a detection model small enough for factory hardware. On a public wood defect dataset it reached a mean average precision of 77.5%, with "only a 0.5 percentage point drop in mAP" on the edge device. That figure and the 2024 paper's 0.94 come from different datasets and defect classes, so they can't be compared directly. Neither is a yield figure.

Reading it through a data lens

In our view, the trial to ask for is the new system against the plant's current method, on matched boards, with volume and value recorded for both. The errors worth counting are the ones that change the cut, and false defects belong on that list as much as missed ones. Volume yield and value yield can also diverge. Cutting around every defect maximises clear pieces, but revenue depends on which lengths and grades the plant can sell.

When it doesn't apply

Plants that cut to fixed customer lengths, with no choice of where to cut, gain little from defect-guided cutting. Products that allow defects in the grade need less precise detection. And the evidence is thin: 100 boards in one recent trial, one 1998 test and one public dataset.

Quarri for wood products is built around how a wood products plant runs, where the order book meets real capacity.

Quarri is an AI-native data platform for the timber supply chain. It connects buying, production, sales and inventory for forest management, sawmill, wood products and pulp, paper and packaging operators.